The Death of the Job: How AI and Robots Will Rewrite Work in the Next 10 Years

Posted on August 20, 2026 by Simon Roses

Read Time: 16 minutes

TL;DR

Work is not disappearing. The job is. The salaried, 9-to-5, one-employer-for-decades package we call “a job” is an industrial-age artifact, and AI agents plus humanoid robots are dismantling it piece by piece. Over the next 10 years, repetitive, risky, and simple tasks get automated at both ends: cognitive work by agents, physical work by robots. The 9-to-5 schedule dies first. Remote and hybrid work keep growing even as corporations push back with return-to-office mandates. Companies stop hiring employees and start engaging experts on missions, a contractor model measured in projects, not years of tenure. Universities that keep selling four-year degrees for jobs that won’t exist must reinvent themselves or become irrelevant. The winners of this transition will be the people who treat AI as an amplifier, build a public reputation, and never stop learning. The losers will be the institutions that pretend nothing is changing.

The Job Is an Industrial-Age Artifact

Let’s start with an uncomfortable truth: the “job” as we know it is not a law of nature. It’s a technology. It was invented.

The 9-to-5 schedule, the salaried contract, the single employer, the office, the 40-year career capped with a retirement watch: all of it was designed for the factory and the corporation of the 20th century. Synchronized hours made sense when work meant standing next to a machine or pushing paper through a hierarchy. You needed everyone in the same place at the same time because coordination was expensive.

Coordination is no longer expensive. Intelligence is no longer scarce. And physical labor is about to stop being exclusively human.

I’ve spent the last two years working alongside AI agents daily, as I described in My Experience Using OpenClaw. My agent works while I sleep. It doesn’t have a schedule. It doesn’t have an office. It doesn’t have a job title. And increasingly, neither will we.

Here is my thesis for the next decade: work survives, the job doesn’t. What replaces it is smaller, faster, more fluid, and much more demanding of the one thing machines don’t have: human judgment.

The Automation Wave Is Real This Time

Every automation panic in history ended the same way: more jobs, not fewer. Economists love to point this out. But “eventually more jobs” and “your job survives” are two very different statements, and the transition is where careers go to die.

I’ll steelman the other side, because it has the historical record behind it. The loom, the tractor, the spreadsheet: each provoked exactly this panic, and each time the economy invented more work than it destroyed. Economists even have a name for assuming otherwise, the lump-of-labour fallacy, the error of treating the amount of work in the world as fixed. Serious people make the optimistic case today too; MIT’s David Autor argues AI could rebuild middle-class work rather than hollow it out, by putting expert judgment back into more hands. That case may well be right. But notice it is an argument about the destination, not the journey, and I am writing about the journey: the five to ten years in which the tasks vanish faster than the institutions adapt. You can believe the long-run optimists and still be the one whose footing goes out from under them in the meantime.

The numbers say this wave is structural, not hype. The World Economic Forum’s Future of Jobs Report projects 170 million new roles created and 92 million displaced by 2030: a net gain of 78 million, but 22% of all jobs churned in five years. Nearly 40% of the skills required on the job will change. And 41% of employers openly plan to reduce headcount as AI automates tasks.

The displacement is not evenly distributed, and this is the part that should worry you. Stanford’s Erik Brynjolfsson and his team, using payroll data from millions of workers, found what they call the “canaries in the coal mine”: employment for workers aged 22 to 25 in AI-exposed occupations is falling at 3.8% per year as of April 2026, while the same age group in low-exposure roles grows at 2%. By the Stanford Digital Economy Lab’s mid-2026 update, that young, exposed cohort sits roughly 19% below where it would be had it simply tracked its less-exposed peers. The tasks vanishing first are exactly what you’d expect: information retrieval, summarization, scheduling, formatting, mechanical assembly of documents. The bottom rungs of the career ladder are being sawed off.

Which tasks get automated follows a simple pattern, and it’s the one I listed years ago for security automation: repetitive, risky, and simple. If your daily work is predictable enough to describe in a prompt, an agent will do it. If it’s dangerous enough to require hazard pay, a robot will do it. If it’s simple enough to learn in a week, software already does it.

As I argued in AI Must Make Superhumans, Not Unemployed, companies that respond to this with mass layoffs are showing a failure of imagination, not a mastery of technology. But my opinion doesn’t change the math: the tasks are going, whether leadership uses the freed capacity to do more or to employ fewer.

Robots Take the Physical Half

For decades, automation was a white-collar spectator sport: software ate the office while the warehouse stayed human. That asymmetry is ending.

The humanoid robots are no longer demos. Figure’s robots completed an 11-month pilot at BMW’s Spartanburg plant, loading more than 90,000 sheet metal parts into welding fixtures across 10-hour shifts on a production line that built over 30,000 vehicles — against a target of 99% placement accuracy per shift, though Figure never published how close they actually came. Agility Robotics’ Digit has logged more than 65,000 operating hours across customer sites including GXO, Schaeffler, and Toyota, and its Oregon factory is designed to build up to 10,000 units per year. At the low end, Unitree shipped roughly 5,500 humanoids in 2025 at prices starting around $16,000, one-tenth of Western platforms.

Let me be honest, because the hype cuts both ways: most humanoid programs are still pilots, cycle times are slower than humans, and Tesla’s Optimus, the most famous of them all, is by Musk’s own admission not yet working in factories “in a material way.” We are in the Apple II era of humanoids, not the iPhone era.

But that’s exactly the point. The Apple II era lasted about a decade. A $16,000 robot that works 24/7 without injuries, sick leave, or turnover doesn’t need to be better than a human worker. It needs to be a fraction as good at a fraction of the cost, doing the dull, dirty, and dangerous jobs no one wants: night-shift logistics, hazardous inspection, repetitive assembly. Those jobs go first, and within 10 years the economics become impossible to ignore for everything from construction to elder care logistics. Do the arithmetic and it stops being abstract: a $16,000 machine amortized over three years of two-shift work is a couple of dollars an hour in hardware plus electricity, and it never files a grievance, calls in sick, or gets hurt. Even a $150,000 Western platform drops below high-wage human labour once it runs enough hours. The honest caveats are uptime, maintenance, and the fact that today’s robots still need babysitting, but the direction of that cost curve is not in dispute, and the major bank analysts have all drawn the same line.

The combination is what matters. AI agents automate the cognitive-repetitive. Robots automate the physical-repetitive. What’s left in the middle is the human core: judgment, accountability, creativity, relationships, and taste.

srf_futureofwork_automation_pincer

The automation pincer: agents eat the cognitive-repetitive, robots eat the physical-repetitive, and the human core in the middle — judgment, accountability, creativity, relationships, taste — is what neither side reaches.

The 9-to-5 Is Already Dead

The 9-to-5 assumed something that is no longer true: that your output was proportional to your hours in a chair.

My agent doesn’t keep office hours. It triages my email before I wake up, runs security scans overnight, and drafts code while I’m at dinner with my family. When part of your workforce operates 24/7, measuring the human part in synchronized 8-hour blocks is absurd. The unit of work is shifting from hours to outcomes: this feature shipped, this audit delivered, this client problem solved.

You can see the schedule cracking everywhere. Four-day-week trials keep expanding, and the World Economic Forum notes that AI-driven productivity is the argument making it viable — organizations that fold AI into redesigned processes can bank the time savings as a shorter week instead of just more output. Hand a knowledge worker back a day’s worth of grunt work and the fifth day is already paid for. Asynchronous work, compressed weeks, project sprints followed by real rest: these are not perks anymore, they are the operating model that matches how augmented humans actually produce value.

Within 10 years, I expect “what are your working hours?” to sound as antiquated as “which fax number should I use?”. You will be paid for judgment and results, and judgment doesn’t punch a clock.

Remote Work: Not for Everybody, But Unstoppable

Here the data looks contradictory, and it’s worth reading carefully because both sides are real.

On paper, the office is winning: by mid-2026, 87% of new US job postings are fully on-site, with just 3% fully remote, as return-to-office mandates pile up. But the workforce hasn’t moved with them: 46% of professionals are already looking or planning to look for a new job, and flexibility is a top reason why — 64% say work-life balance and remote options would make them switch employers. Companies are mandating a model their own talent is quietly heading for the door to escape.

My prediction: the mandates lose, slowly, and for a cold economic reason. When you hire an expert for a mission instead of an employee for a desk (more on that below), geography stops mattering. The best AI security specialist for your project might be in Madrid, Bangalore, or São Paulo, and she is not relocating for a six-month engagement. Companies that insist on presence will select from the shrinking pool of people willing to commute; companies that master distributed work will select from the planet.

But let me be equally honest about the other half: remote work is not for everybody, and pretending otherwise has hurt people. It demands self-discipline, written communication skills, a home where deep work is possible, and a personality that doesn’t wither without hallway conversations. Juniors especially suffer: the Stanford data shows their ladder is already being automated away, and remote isolation makes learning-by-osmosis even harder. The future is not “everyone remote.” It’s remote as a skill you deliberately build, hybrid as the default equilibrium, and physical presence reserved for what it’s actually good at: trust-building, mentoring, and creative collision.

Balance Stops Being a Perk and Becomes Infrastructure

Here’s a second-order effect almost nobody prices in: when AI removes the repetitive 60% of your work, what remains is the hard 40%: decisions, creativity, responsibility. That work is cognitively expensive. You cannot do eight hours of pure judgment a day, no human can.

The industrial job diluted hard thinking with meetings, forms, and busywork. The AI-era “job” is concentrated: shorter, denser, heavier per hour. Which means recovery is no longer a lifestyle preference, it’s maintenance of the production asset, and the asset is your mind. Athletes figured this out decades ago: they don’t train 8 hours a day, and nobody calls them lazy.

Companies will learn, some the hard way, that burning out judgment-workers is like redlining an engine: you get one great quarter and then a blown machine. Within the decade I expect work-life balance, real balance, not a wellness app and a pizza Friday, to move from HR brochure to contract clause. Experts negotiating project engagements will price their recovery time in, the same way consultants already price travel. The companies that respect it will get the best people. The ones that don’t will get the people nobody else wanted.

Experts, Not Employees: The Mission Model

This is the biggest structural change of the decade, and the least discussed.

The traditional employment deal was: you give me 40 years, I give you stability, training, and a pension. That deal is already dead; companies just haven’t updated the paperwork. Average tenure keeps falling — US median job tenure slid to 3.9 years in 2024, its lowest since 2002 — “stability” evaporated with every AI-justified layoff round, and loyalty is a one-way street corporations drive trucks down.

What replaces it is the model Hollywood has used for a century: assemble experts around a mission, execute, disband. You don’t “hire an employee.” You engage a specialist, for a project, for as long as the mission lasts: six months, two years, five years. Then everyone moves to the next production.

The numbers show it’s already happening. 72.9 million Americans worked independently in 2025, with the $100K+ earners among them growing 19% in a single year to 5.6 million. And the demand side is moving to meet them: in one survey of tech leaders, 92% expect to increase their engagements with freelance or fractional talent over the next two years. Read that again: the direction of travel is not toward more generic full-time staff. It’s toward fewer, better, temporary experts, because the generic work is exactly what the agents absorbed.

Why does AI accelerate this? Because an expert with agents is a complete unit of production. I run VULNEX with AI leverage that would have required a team of ten a few years ago. The expert brings judgment and reputation; the agents bring scale. A company no longer needs to warehouse full-time generalists “just in case” when it can plug in a proven specialist who arrives with her own AI infrastructure and delivers from day one.

The consequences cut deep, and not all of them are pleasant:

Your reputation becomes your CV. In a mission economy, you are hired for what you can demonstrably do, not for titles you held. Public work: code, writing, talks, tools, compounds into the asset that gets you the next mission. Invisible excellence stops paying.

The safety net breaks. Health insurance, pensions, sick leave, mortgage eligibility: entire social systems assume the employee contract. A workforce of mission-based experts needs portable benefits, and governments are a decade behind. The serious proposals already exist — portable benefit accounts that follow the worker, sectoral funds, wage insurance, the perennial universal-basic-income debate — but most reskilling programs today are theatre, and pretending a laid-off logistics worker becomes a prompt engineer is the same delusion in a nicer suit. This will be one of the defining political fights of the 2030s, and countries that solve portable protection first will attract the world’s best independent talent.

Not everyone is built for it. The mission model rewards self-starters with rare skills and punishes people who need structure. If we’re honest, the old job was also a social technology for giving ordinary people stable lives. Its death creates real losers, and pretending everyone can be a personal brand is Silicon Valley delusion. Society will need answers here that go beyond “learn to freelance.”

The Security Bill Nobody Is Costing

Now I’ll put my other hat on, because almost nobody debating the future of work looks at it from a security chair, and the mission economy is a security problem wearing an HR costume.

Think about what “fewer employees, more experts on missions” does to your attack surface. Every full-timer you swap for a rotating cast of specialists is an identity to provision and — the part everyone forgets — to deprovision. Access that used to sit inside a badge and a managed laptop now sprawls across contractors’ own devices, their own cloud tenants, their own AI tools. Intellectual property walks in and out with every engagement. The insider threat is no longer a disgruntled lifer; it’s a stranger with legitimate access for ninety days and no reason to protect you after. Freelance-marketplace accounts get phished and resold. And the “own AI infrastructure” that makes an expert a complete unit of production is, from the defender’s side, unmanaged shadow AI touching your data with logging you don’t control.

Then add the robots and agents themselves. A humanoid on the factory floor is an OT/IoT device with cameras, microphones, network access and physical actuators — an attack surface that can now walk. An autonomous agent holding credentials is a privileged account that acts on its own initiative, which is exactly the risk I keep circling in When the Model Is the Attacker. The workforce of 2035 is part human, part agent, part machine, and every one of those parts is something an adversary can target, impersonate, or turn.

None of this is a reason to stop. It’s a reason to build the security model before the org chart dissolves, not after the first breach traces back to a contractor who left six months ago. The companies that win the mission economy will be the ones that treat identity, data governance, and endpoint trust as the foundation of the model, not the paperwork they mean to get to later.

Europe Will Not Live This the American Way

Almost every number above is US data, and Europe will go through this differently — in both directions.

On one side, the friction here is real. European labour law was written to protect the employee, not the mission: strong dismissal protection, works councils, and, in Spain, the famously heavy autónomo regime make “assemble, execute, disband” slower and costlier than it is in Austin or Bangalore. No European employer is churning 22% of its workforce in five years the way an at-will US market can. The transition arrives later here, and more mediated — negotiated through unions and ministries rather than a spreadsheet.

On the other side, the exposure is sharper where it lands. Spain already runs a youth unemployment rate around 23% — roughly one in four under-25s, among the highest in the EU — and the “canaries” data says the entry-level rungs are exactly what AI removes first. A generation that already struggles to get onto the ladder now watches the bottom of it being automated. And the rules are uniquely European: hiring, firing, and worker-management AI are classified as high-risk under the EU AI Act (Annex III), so the same automation reshaping work on the continent arrives wrapped in compliance duties the US never imposes. The mission economy is coming to Europe too. It just has to negotiate with a continent that wrote its labour rules for the world the job built.

Universities Are Selling Maps of a World That No Longer Exists

Now for the institution least prepared for all of the above.

The university’s implicit promise, four years, one degree, one stable career, is collapsing in real time. Computer science, the “safe” degree of the last 20 years, saw enrollment drop 8.1% in the 2025-26 academic year, the steepest fall of any field, with pure CS down 11.2%. Recent CS graduates are now more likely to be unemployed than history and liberal arts majors. Students watched AI write code and did the math their advisors wouldn’t.

The problem is not that education is obsolete. It’s that the methodology is. Universities still optimize for knowledge transfer, lectures, memorization, exams, in a world where knowledge is free and instantly accessible to anyone with an agent. What’s scarce is everything the lecture hall doesn’t teach: judgment under uncertainty, taste, working with AI tools, shipping real things, and learning how to learn continuously.

If I were redesigning a university for the next decade, and they need redesigning, not tuning, I’d change four things:

From degrees to apprenticeships. Medicine got this right centuries ago: you learn by doing, supervised, on real cases. Every discipline needs its residency. A student who has shipped three real projects with AI tools is worth more than one who memorized the textbook the AI already read.

From four years to lifelong subscription. With 39% of skills changing every five years, front-loading education into ages 18-22 is engineering malpractice. The university of 2036 is a place you return to every few years for intensive re-tooling, an institution you subscribe to for a career, not a campus you graduate from once.

Teach judgment, not syntax. Stop teaching what AI does well. Teach what it does badly: framing problems, questioning outputs, ethics, security thinking, first principles. I made the same argument for developers in Professional Vibe Coding vs. Vibe Coding: the value is no longer typing the code, it’s knowing when the machine is wrong.

Make AI fluency the new literacy. Every graduate, philosopher or physicist, should leave knowing how to direct agents, verify their output, and secure them. A university that bans AI tools in 2026 is a swimming school that bans water.

The universities that adapt will thrive, because the demand for learning has never been higher. The ones that keep selling the old map will follow the fate of every institution that mistook its format for its mission.

What Should You Do? My Practical Bets

I’m a security guy; I don’t do predictions without mitigations. If the next 10 years look anything like the picture above, here is the personal playbook:

  1. Become AI-native now. Not “I tried ChatGPT once.” Agents doing real work in your daily workflow. The gap between AI-augmented professionals and everyone else is compounding monthly, and it’s already visible in output.
  2. Move up the judgment stack. Audit your own tasks: everything repetitive, predictable, or simple in your role is on the automation menu. Deliberately migrate your value toward decisions, architecture, relationships, and accountability, the things someone must still sign their name to.
  3. Build in public. Your next mission will come from your visible track record, not from an HR keyword filter. Write, publish tools, give talks, show your work. Reputation is the currency of the expert economy, and it compounds like interest.
  4. Structure yourself as an expert, even while employed. Treat your current job as a mission among missions: keep your skills liquid, your network warm, and your finances able to survive gaps between engagements. Employment is no longer a pension plan; it’s a client.
  5. Own your infrastructure. As I wrote after Anthropic locked subscriptions out of third-party agents: dependence on any single provider, employer or AI vendor, is a vulnerability. Local models, your own tools, your own audience. Sovereignty scales down to individuals.
  6. Guard your recovery like a deliverable. Judgment is your product and it degrades with exhaustion. Schedule recovery with the same seriousness you schedule delivery. Nobody else will do it for you, least of all the mission economy.

And If You’re the One Leading

That list is for the individual. If you run a company or a security team, the same decade lands on your desk as a different set of choices, and I’ll be just as blunt.

Don’t confuse a layoff with a strategy. As I argued in AI Must Make Superhumans, Not Unemployed, cutting heads because an agent absorbed a few tasks is the lazy move. The freed capacity is a chance to do more with the judgment you already pay for, not an excuse to have less of it.

Retain judgment, rent scale. Keep the people who own decisions, accountability, and relationships on the inside. Bring in mission experts for the spikes. And budget the recovery of the judgment-workers you keep, because you are running an expensive engine and burnout is how you throw a rod.

Make security a precondition, not a cleanup. If you’re going to run on contractors, agents, and robots, the identity, data-governance, and endpoint story has to exist first — not after the post-mortem. See the section above; it is cheaper as architecture than as incident response.

Own your dependencies. Your leverage, and your risk, increasingly sit with a handful of AI vendors. Engineer for portability the way you’d refuse a single-supplier lock-in anywhere else in the business.

The Bottom Line

The job, the 9-to-5, single-employer, salaried package, was a brilliant technology for the industrial age, and it’s reaching end-of-life. AI agents are absorbing repetitive cognitive work at a pace measured in months. Humanoid robots are leaving the demo reel and clocking real factory hours. The schedule is dissolving into outcomes, the office into networks, and the employee into the expert-on-mission.

None of this means the end of work. The WEF math still comes out positive: more roles created than destroyed. But the transition will be brutal for everyone who assumes their job description is a load-bearing wall, for the juniors whose ladder is being automated away, and for institutions, universities first among them, that keep selling stability they can no longer deliver.

Ten years from now, the people thriving won’t be the ones who competed with the machines, or the ones who ignored them. They’ll be the ones who did what humans have always done with a new tool: picked it up, mastered it, and used it to do work no machine, and no un-augmented human, could do alone.

The job is dying. Long live the work.

Further Reading:

Posted in AI, Business, Technology | Tagged , | Leave a comment

When a Missing Patch Becomes a Defective Product: The New EU Product Liability Directive

Posted on August 14, 2026 by Simon Roses

Read Time: 13 minutes

TL;DR

On December 9, 2026 — a few months from now — the new EU Product Liability Directive ((EU) 2024/2853, “PLD”) starts applying to products placed on the EU market. For the first time, software is explicitly a product under strict liability rules: standalone apps, firmware, SaaS, AI systems, even digital manufacturing files. A product is legally defective if it lacks the cybersecurity a person is entitled to expect — and the directive says out loud that failing to ship security updates for a vulnerability under your control can make your product defective. Compensable damage now includes destruction or corruption of your personal, non-professional data and medically certified psychological harm, the old €500 threshold is gone, courts can order you to disclose technical evidence, defectiveness gets presumed when cases are too technically complex for the claimant, and you cannot disclaim any of this in your EULA. There is no US equivalent: the US still runs product liability on state tort law where “is software a product?” remains unresolved, and the 2026 US Cyber Strategy explicitly walked back the previous administration’s plan to shift liability onto software makers. The two biggest software markets on Earth are now driving in opposite directions — and if you sell into the EU, your vulnerability management program just became your legal defense file. Here’s my security-practitioner read.

The usual disclaimer, doubled: I am not a lawyer, and this is not legal advice — it’s a security practitioner reading a liability regime the way an attacker reads a network diagram, looking for where the pressure actually lands. If you ship software into the EU, talk to actual counsel. What I can tell you is what this changes operationally for the people who build and secure products, because I have spent twenty-plus years watching the industry treat security as a best-effort promise backed by a liability disclaimer. That era has an expiry date in Europe, and the date is December 9, 2026.

I have been circling this theme on the blog for a while — who owns the risk when AI writes your code, why companies have no AI strategy but plenty of AI exposure, what happens when nobody is accountable for an agent’s actions. The PLD is the EU answering a chunk of those questions with a blunt instrument: someone in the supply chain is always liable, and it is never the victim.

This post has a companion. A few weeks after I first drafted it, the regulatory half of Europe’s push went live: the AI Act’s GPAI enforcement, in force since August 2. A regulator that fines you and a courtroom that bills you are two different doors — that post is the regulator; this one is the courtroom.

What Changed: From 1985 to 2024

The old Product Liability Directive dates from 1985 — the year of the Amiga and the C64 in its prime. It was written for toasters and lawnmowers: physical products, physical harm. Software mostly escaped it. Whether code was even a “product” was debated for four decades, pure data loss wasn’t compensable damage, there was a €500 floor on property claims, and proving that a specific defect in a complex system caused your harm was on you, alone, against the manufacturer’s information advantage.

Directive (EU) 2024/2853 replaces all of that for products placed on the market from December 9, 2026 (products shipped before that date stay under the 1985 rules — a long-tail dual regime worth understanding). Member states must transpose it into national law by that same date, and they are predictably behind: as of mid-2026, Hungary has transposed, a handful of countries (Croatia, Slovakia, Bulgaria) have drafts moving, and much of the EU has published nothing yet. Late transposition won’t save anyone — the deadline and the direction are fixed.

The headline changes, from a builder’s chair:

Software is a product. Full stop. Standalone software, embedded firmware, mobile apps, AI systems, SaaS and cloud-delivered functionality, related digital services integrated into a product (think the health-monitoring service behind a wearable’s sensors), and even digital manufacturing files (the CAD file that 3D-prints the part). If you place it on the EU market, strict liability attaches — no fault, no negligence required. The claimant needs defect, damage, and causal link. Not your intent, not your process, not your apology.

Someone is always on the hook. Liability cascades: the manufacturer (which includes the software developer) first; for non-EU manufacturers, the importer or authorized representative; failing that, the fulfilment service provider; even distributors and online marketplaces if they can’t identify who’s upstream within a month. The EU has deliberately closed the “the vendor is a shell in another jurisdiction” escape hatch. If you develop from outside the EU and sell in, your EU representative is holding your liability.

02-liability-cascade-graph

Figure 1. The cascade: liability flows manufacturer → importer / authorised representative → fulfilment provider → distributor / marketplace until it lands on someone reachable in the EU. Non-EU shells don’t break the chain, and none of it can be disclaimed.

You can’t contract out. Liability toward injured persons cannot be excluded or limited. The all-caps AS IS, WITHOUT WARRANTY OF ANY KIND block at the bottom of the license? In this regime, decorative.

The Security Part: Defectiveness Now Speaks CVE

This is the section that made me write this post. The directive doesn’t treat security as an add-on; it bakes it into the legal definition of “defective.”

A product is defective when it does not provide the safety a person is entitled to expect, and the assessment explicitly includes, among other factors, relevant cybersecurity requirements and the product’s ability to withstand foreseeable third-party actions — i.e., attacks. Read that again: a foreseeable attack exploiting a weakness you should have addressed is part of the defect analysis. The exploit doesn’t excuse you; its foreseeability indicts you.

01-defectiveness-tree

Figure 2. What a claimant has to line up for strict liability — defect AND damage AND causation — and the security failures that make software “defective”: missing cybersecurity, unshipped updates, a foreseeable exploited weakness, or CRA/NIS2 non-compliance. Damage now includes destroyed data and psychological harm; causation can be presumed.

Three consequences I’d put on any product team’s wall:

1. Unpatched vulnerabilities are now a legal category, not just a backlog item. The directive holds manufacturers responsible for products after they ship, to the extent the product stays within their control — software updates, upgrades, cloud-side behavior, and machine-learning-driven change all count. Failing to supply the security updates needed to maintain safety can itself make the product defective. The mental model shifts from “we ship, users assume the risk” to “the product carries an ongoing duty of care as long as we can push code to it.”

2. Your compliance posture feeds the defect analysis. The PLD interlocks with the Cyber Resilience Act, NIS2, and sectoral rules like the Medical Devices Regulation. Non-compliance with mandatory security requirements doesn’t just bring regulatory fines anymore — it becomes evidence of defectiveness in a private damages claim, and breach of mandatory safety requirements can trigger a presumption of defect. The CRA tells you how to build and maintain; the PLD is what a claimant beats you with when you didn’t. Regulatory compliance and civil liability used to be separate lanes. They just merged.

3. The development-risk defense has a software-shaped hole. Manufacturers keep the classic “state of scientific and technical knowledge” defense — the defect was unknowable when we shipped. But it does not rescue you for defects that emerge from software updates or evolving ML behavior under your control after shipment. And “substantial modification” of a product — which a major update can be — restarts the liability clock. For a continuously-deployed SaaS with a learning model in the loop, the 10-year longstop is less a finish line and more a treadmill.

The Privacy Part: Data Loss Is Now Damage

Here is the sleeper provision. Compensable damage under the new PLD includes the destruction or corruption of data not used for professional purposes. Alongside death, personal injury — now including medically certified psychological harm — and property damage, with the €500 threshold deleted.

Think about what that covers in practice:

  • Ransomware tears through a consumer NAS because of a known, unpatched vulnerability in its firmware → the family photos it encrypted are compensable damage in a strict-liability claim against the manufacturer.
  • A defective sync client corrupts a decade of personal documents → damage.
  • An IoT hub update bricks devices and wipes local data → damage.

And this stacks alongside GDPR, it doesn’t replace it. GDPR Article 82 already gives compensation for unlawful processing; the PLD adds a parallel track where the claim isn’t “you processed my data unlawfully” but “your defective product destroyed my data.” Different defendant theory, no need to prove a GDPR infringement, strict liability instead of the controller-accountability dance — and consumer organizations can bring these as representative (collective) actions. A widespread incident caused by a negligent-patch-cadence product is no longer just a PR problem and a possible DPA fine; it’s a class-action-shaped liability with per-user damages that now include the data itself.

The evidentiary machinery makes it bite. Courts can order disclosure of technical documentation and evidence — logs, risk assessments, internal vuln reports — presented in an accessible form; refuse, and defectiveness is presumed. And when the claimant faces excessive difficulty proving defect or causation due to technical or scientific complexity — which describes essentially every AI system and most distributed software — courts can presume those elements too. The information asymmetry that quietly protected software vendors for forty years was a design flaw, and the EU patched it.

One more wrinkle for my corner of the industry: the AI Liability Directive is dead — the Commission announced its withdrawal in early 2025 and formally scrapped it later that year. So for AI systems in the EU, the liability stack in 2026 is exactly this PLD (AI systems are software, software is a product) plus the AI Act’s regulatory duties plus national fault-based rules. The PLD’s complexity presumption was practically written with “explain your model’s decision chain to a judge” in mind.

Is There a US Equivalent? No — and It’s Getting Less Equivalent

Short answer: no. Longer answer: the US has product liability, but nothing like this, and the gap is widening on purpose.

US product liability is state tort law — strict liability, negligence, and warranty theories, shaped by the Restatement of Torts §402A and a patchwork of fifty state variations. There is no federal product liability statute, no EU-style harmonized regime. And critically for us:

  • Whether software is a “product” at all is unresolved. The Restatement’s strict liability tradition covers tangible goods; courts remain split on apps, platforms, and algorithms, and vendors argue software is a service precisely to stay out of products-liability land. The EU spent 2024 settling by statute the question US courts are still litigating case by case.
  • The economic loss doctrine blocks most of what the PLD just opened. In most states, if a defective product only damages itself or causes purely economic/data harm, tort recovery is barred — you’re pushed into contract, where the EULA is waiting.
  • And the EULA works. US software licensing lives on disclaimers and liability caps that are broadly enforceable. The exact instrument the PLD voids is the load-bearing wall of the US software industry’s risk model.

What the US does have is a patchwork of pressure: FTC Section 5 enforcement against companies with sloppy security, state IoT security laws like California’s SB-327, FDA premarket cybersecurity for medical devices, the voluntary Cyber Trust Mark label. Real, but scattered, mostly regulatory rather than private-liability, and nothing that hands a consumer a strict-liability claim for a destroyed hard drive.

The trajectory is the interesting part. The 2023 US National Cybersecurity Strategy (Pillar 3) proposed exactly the EU move: shift liability onto software makers that ship insecure products, paired with a safe harbor for those who demonstrably develop securely. It was the boldest software-liability language ever to come out of the White House — and it was never legislated. The March 2026 Cyber Strategy for America then explicitly departed from it: deregulation, “cyber defense should not be reduced to a costly checklist,” compliance-burden reduction, no liability shift, no safe harbor.

So as of mid-2026 the divergence is official policy on both sides: the EU made insecure software a defective product; the US decided not to. If you ship globally, you will build to the EU bar anyway — the same Brussels-effect logic as GDPR — because maintaining two security postures is more expensive than maintaining one. Which means the PLD is de facto setting the global floor for software liability, from Brussels, without the US Congress ever voting on it.

Open Source: Mostly Safe, With a Commercial Tripwire

The directive excludes free and open-source software developed or supplied outside the course of a commercial activity. The hobbyist maintainer, the academic project, the community library on GitHub — out of scope, and rightly so; strict liability on unpaid maintainers would have been an extinction event for the commons.

But the tripwire is the word commercial. Charge for the software, sell support around it, or monetize personal data beyond what’s needed for security/compatibility, and the exemption evaporates. And the moment an OSS component is integrated into a commercial product, the integrator owns the liability for it. Your product’s SBOM is now a liability map: every dependency in it is a component you answer for in an EU courtroom. I wrote about the dependency trap in AI-generated code — the PLD is that post with a court attached. “The vulnerability was in an upstream library” has never been much of an excuse technically; from December it isn’t one legally either.

My Read: What Security Teams Should Actually Do

Strip the legalese and the PLD is a list of operational demands. Most of them are things good security teams already preach; the difference is that “we didn’t get around to it” now has a price tag with a court attached.

1. Your vulnerability management program is now your legal defense file. SLA-driven patching, documented triage decisions, EOL and support-window policy, coordinated disclosure handling — these stop being maturity-model line items and become the evidence you’ll produce under a disclosure order. Run vuln management assuming every decision may be read aloud to a judge, because under the disclosure regime, it may. If you can’t show why you deprioritized the bug that later destroyed someone’s data, the presumption machinery does the rest.

2. Update capability is a liability boundary. “Within the manufacturer’s control” is the load-bearing phrase of this directive. If you can push updates, you carry the duty; how long you promise to is your support-lifecycle policy. Define it, publish it, honor it, and treat end-of-support as a formal, dated, communicated event — not a quiet cessation of patches. And notice the perverse incentive worth designing around: teams may be tempted to reduce update capability to shrink the control window. Wrong answer — the CRA independently mandates security support. The only way out is through.

3. Logs are exculpatory evidence now. The same telemetry I keep demanding for agent security does double duty here: proving what your product did, when you knew, and how fast you moved is how you rebut a defect presumption. A product that can’t reconstruct its own behavior can’t defend itself — in the SOC or in court.

4. Compliance artifacts are double-edged — keep them honest. CRA conformity work, risk assessments, threat models: they’re your defense when real, and a claimant’s Exhibit A when aspirational. A threat model that lists a risk you then demonstrably ignored is worse than no document at all. Write what you’ll do; do what you wrote.

5. If you’re a US vendor, don’t relax. Your home regulator just stood down, but your EU importer or authorized representative is holding your strict liability, and they will pass that risk back to you through contracts, audits, and insurance requirements. The commercial chain will transmit the PLD’s pressure across the Atlantic faster than any regulation would have.

So What

For forty years, software has been the only mass-market engineering discipline allowed to ship known-defective products to the public and disclaim the consequences in a click-through. Bridges can’t do it, cars can’t do it, toasters can’t do it. From December 9, 2026, in the EU, software can’t do it either.

The cynical read is “more European red tape.” I don’t buy it, and you know I’m no reflexive fan of regulation. Strict liability is not a checklist — it’s the opposite of one. It doesn’t tell you how to build; it tells you that if what you build hurts someone, you pay, and lets you engineer your own way to that bar. That’s the deal every other engineering field has operated under for a century, and those fields responded by inventing safety engineering, not by collapsing.

Meanwhile the US just bet the other way: that market incentives without liability will produce secure software. We’ve run that experiment for four decades. The result is the CVE list.

The vendors who treated security as engineering all along have little to fear from December. The vendors who treated it as a disclaimer are about to discover their EULA was never load-bearing. Patch like it’s a legal duty — because in Europe, it now is.

Stay paranoid. Ship patches. Keep your logs.

Further Reading:

Questions or feedback? Reach out via:

Need help getting your product ready for the PLD/CRA era? VULNEX offers:

  • Product & application security assessments (secure-by-design gap analysis, SBOM and supply-chain review)
  • Vulnerability management program design (SLA-driven patching, disclosure handling, support-lifecycle policy)
  • AI system security assessments and liability-aware threat modeling
  • Red team engagements and security automation

For AI security strategy — where model and agent risk meets board-level decisions — see vulnex.ai.

Contact: info@vulnex.com

Posted in AI, Security, Technology | Tagged , , , , , | Leave a comment

The Day the AI Act Grew Teeth: GPAI Enforcement Goes Live

Posted on August 8, 2026 by Simon Roses

Read Time: 13 minutes

TL;DR

On August 2, 2026, the part of the EU AI Act everyone was quietly ignoring became enforceable: the AI Office can now fine providers of general-purpose AI models up to 3% of global annual turnover or €15 million, whichever is higher, and can demand your technical documentation, run its own evaluations of your model, order you to “take measures,” and in the worst case make you restrict, withdraw, or recall the model from the EU market. The obligations themselves have technically existed since August 2, 2025 — technical documentation, downstream transparency, a copyright policy that honors robots.txt and opt-outs, a public summary of training data — but until now they were rules without a referee. That changed. Models judged to carry systemic risk (trained above 10^25 FLOP) carry heavier duties that read like a security checklist written by a regulator: model evaluations, adversarial testing / red-teaming, a safety-and-security framework, serious-incident reporting to the AI Office, and cybersecurity protection of the model weights themselves. The voluntary Code of Practice buys you a lighter touch, not immunity. Models already on the market before August 2, 2025 get until August 2, 2027 to fall in line. This is the regulatory half of a two-part story — the civil-liability half, the new Product Liability Directive, is the post I’m publishing right after this one. Read together, the EU has built a pincer: a regulator that fines you, and a courtroom that bills you. Here’s my security-practitioner read of the regulatory jaw.

The dates that matter:

Date What happens
Aug 1, 2024 The AI Act enters into force
Aug 2, 2025 GPAI model obligations begin (for models placed on the market after this date)
Aug 2, 2026 Enforcement switches on — the AI Office / Commission can investigate, evaluate, order measures, and fine
Aug 2, 2027 Compliance deadline for GPAI models already on the market before Aug 2, 2025

The usual disclaimer, same as always: I am not a lawyer, and this is not legal advice — it’s a security practitioner reading a regulation the way I read an attack surface, looking for where the pressure actually lands and who ends up holding it. If you build or ship AI models into the EU, talk to actual counsel. What I can tell you is what this changes operationally for the people who build, secure, and deploy these models — because buried under the compliance language is a list of things I have been demanding on this blog for two years, now backed by a fine.

A note on framing before we start. This post is one of a pair. The EU is putting teeth behind AI on two different tracks at once, and they bite in different ways. This one — the AI Act’s rules for general-purpose AI — is regulatory: a public authority, the AI Office, with the power to investigate you and fine you. The companion piece, on the new Product Liability Directive, is civil: private plaintiffs and courts, strict liability, damages paid to the person your defective software harmed. I’m publishing them back to back on purpose, because if you only track one you’ll misjudge your exposure. A regulator fining you and a claimant suing you are two separate doors, and after this summer both are open.

I have been circling the regulatory door for a while — why “we use ChatGPT” isn’t an AI strategy, what happens when the model itself becomes the attacker, whether you can still tell an open-weight model from a frontier one. August 2 is the EU answering a slice of those questions with an enforcement budget attached.

What Actually Changed on August 2

Here’s the part that confuses people, so let me be precise: August 2, 2026 did not create new obligations. The substantive rules for general-purpose AI (GPAI) models kicked in a full year earlier, on August 2, 2025. What was missing until now was the enforcement machinery. For twelve months the AI Act’s GPAI chapter has been law you could technically break without anyone able to do much about it.

That grace period is over. As of August 2, 2026, the AI Office, the Commission’s dedicated AI enforcement body, has four concrete powers it did not have on August 1:

  1. Demand your documentation. It can require a GPAI provider to hand over technical documentation and information about the model.
  2. Evaluate your model. It can run its own assessments of your model to check compliance and investigate systemic risk — including requesting access.
  3. Order compliance measures. It can require you to “take appropriate measures” to bring the model into line.
  4. Pull the model. In the worst case it can make you restrict its availability, withdraw it, or recall it from the EU market.

One precision point worth keeping straight, because the lawyers reading this will: the AI Office is the operational body that investigates, evaluates, and builds the case, but the formal decision to fine under Article 101 is the European Commission’s. In practice you deal with the AI Office; the signature on the penalty is the Commission’s.

And behind all four sits the number that focuses minds: fines of up to 3% of global annual turnover or €15 million, whichever is higher, under Article 101. Note that’s the GPAI-specific ceiling; the Act’s headline 7%-of-turnover fines are for deploying prohibited AI practices, a different regime. But for a frontier lab, 3% of global turnover is a board-level number.

01-enforcement-tree

Figure 1. How enforcement actually lands: a non-compliance gap that survives the AI Office’s escalation ladder — documentation request → model evaluation → compliance order — ends in a fine of up to 3% of global turnover or €15M, and at the extreme, restriction or withdrawal from the EU market.

So nothing about your model’s obligations changed this week. What changed is that ignoring them now has a price, a referee, and a stop button.

And here is the tell that this date is real. The EU’s Digital Omnibus — a simplification package the industry lobbied for hard, tabled in November 2025 and agreed this spring — postponed the AI Act’s high-risk deadlines, sliding the Annex III obligations to December 2027 and embedded systems into 2028. It left GPAI alone. Of all the deadlines Brussels was pressed to move, the one it would not move is the one this post is about. When a regulator blinks on nearly everything except the thing you are writing about, that thing is the priority.

Who This Actually Hits

The AI Act is fussy about roles, and the fine print matters. The GPAI obligations land on the provider of the model — the lab that trains and places the general-purpose model on the market. Think the obvious frontier names, but also the growing field of open-weight labs, and, importantly, anyone who fine-tunes or substantially modifies a model to the point of becoming, in effect, its new provider. The Commission’s own guidance puts a rough line on that: modify a model using more than about a third of its original training compute and you’re presumed to have become a provider yourself, with the obligations that follow. That clause is the one that pulls a lot of companies who think of themselves as mere “users” into scope without noticing.

If you’re a deployer — you build a product on top of someone else’s model — most of these GPAI duties are not directly yours yet (your day comes later, once the high-risk-system rules bite — a timeline the Digital Omnibus just pushed back and made conditional on technical standards). But you inherit the consequences: the transparency information your upstream provider must now give you is exactly the material your own compliance, and your own security review, depend on. Which is the first place a security practitioner should perk up: the Act is forcing your model vendor to tell you things they previously treated as trade secrets. Use that.

The Baseline: What Every GPAI Provider Now Owes

For every general-purpose model on the EU market, regardless of size, four duties:

Technical documentation. A detailed, maintained dossier on the model — architecture, training process, intended and excluded uses, energy consumption — retained for ten years and produced to the AI Office on request. This is the file that gets read aloud in an investigation.

Downstream transparency. You must publish contact details and respond to downstream providers’ requests with the information they need to integrate the model responsibly — the Code of Practice’s transparency guidance points at a short, days-long response window (law-firm readings cite roughly 14 days) — while still protecting legitimate IP and trade secrets. The “it’s all proprietary, figure it out yourself” era of model integration is ending.

A copyright policy with actual mechanics. Not a paragraph of intent — a working policy that respects technological protection measures, excludes known piracy sources, honors robots.txt and machine-readable opt-out signals, and gives rightsholders a contact and a complaint path. This is the provision the training-data lawsuits will hang on.

A public training-data summary. A filled-in AI Office template summarizing what the model was trained on. Not the dataset, but enough of a summary that the black box gets a label.

None of this is exotic to anyone who has run a mature engineering shop. What’s new is that it’s mandatory, enforceable, and discoverable.

Systemic Risk: When the Regulation Starts Speaking My Language

Here is the section that made me want to write this post. A subset of models, those with “high-impact capabilities” (presumed once training compute crosses 10^25 FLOP) plus any others the Commission designates, are classified as carrying systemic risk. And the obligations that attach to them could have come straight off one of my own engagement checklists:

  • A safety and security framework (the timing specifics here come from the Code of Practice’s safety-and-security chapter), stood up within weeks of notification and finalized before the model ships.
  • Model evaluations and adversarial testing — the Act says red-teaming out loud. State-of-the-art evaluation of the model’s dangerous capabilities is now a legal duty, not a nice-to-have your safety team fights for budget on.
  • Systemic-risk assessment and mitigation across the lifecycle: filtering, monitoring, input/output controls.
  • Serious-incident reporting to the AI Office and national authorities on staggered timelines — an actual incident-response obligation for models.
  • Cybersecurity protection of the model and its physical infrastructure — i.e., protect the weights. Weight exfiltration is now a compliance failure, not just an embarrassing headline.
  • Ten-year documentation retention, and a safety-and-security report including external evaluators’ findings.

A fair note on sourcing: the Act itself sets these duties at the level of principle (Article 55); several of the concrete specifics above — the framework’s timing, the exact shape of the safety report — come from the Code of Practice’s safety-and-security chapter, which is the paved road for demonstrating you met the statutory bar. The duty is law; some of the detail is the Code.

Read that list and then reread what I wrote after the Hugging Face / OpenAI model-evaluation incident: most defenders have zero telemetry at the model and agent layer, and the tooling to run a full intrusion at machine speed is now something you can trigger by accident during your own safety testing. The AI Act just made the telemetry, the evaluation, and the incident reporting for that exact layer a regulated obligation for systemic-risk models. I don’t love every line of this Act, but I’m not going to pretend that mandating adversarial testing and weight security for the most capable models on Earth is the part worth complaining about. It’s the part I’ve been asking for.

There’s also a subtle security-economics point. Adversarial testing is only as good as the adversary. A regulation that requires red-teaming without defining rigor invites the checkbox version — a friendly internal team, a weekend, a green report. The labs that treat this as real offensive work (external, adaptive, incentivized to actually break the model) will produce safety evidence that means something; the ones that treat it as a compliance artifact will produce a document that looks great right up until someone reads it in an investigation. Same lesson as every compliance regime I’ve ever worked under: the artifact is only worth the honesty behind it.

The Open-Weight Wrinkle

This is where my open-weight models thesis collides with the Act, and it is the most interesting knot in the whole thing. The AI Act gives open-source / open-weight GPAI models a partial break: models released under a truly free and open license, with their parameters and architecture made public, are exempted from some of the baseline obligations — the technical documentation and the downstream-transparency duties — though not the copyright policy or the public training-data summary, which every provider owes regardless. The logic mirrors the open-source carve-out I described in the liability piece: you don’t want to crush the commons under paperwork.

But — and it’s the same but as everywhere else in EU tech law — the exemption evaporates the moment the model carries systemic risk. Cross the compute threshold, and open weights save you from none of the heavy duties: you still owe evaluations, adversarial testing, incident reporting, and weight security. Which lands in a peculiar spot given where the frontier is going. As I wrote last month, Chinese labs are shipping trillion-parameter open models that trade blows with US frontier systems, NVIDIA and Meta are shipping open too, and at VULNEX we already run our own offensive agent on an open-weight model. The Act’s structure means the most capable open models — precisely the ones the sovereignty argument is most excited about — inherit the most regulatory weight. Openness buys you a lighter touch only until your model gets good enough to matter.

For enterprises running open weights on their own hardware, there’s a quieter implication I’ll flag: if you fine-tune an open model far enough, you may become a provider in the Act’s eyes, and inherit obligations you assumed belonged to the lab you downloaded from. “We just run it locally” is not the force field people think it is.

The Code of Practice: A Lighter Touch, Not a Shield

The Commission published a GPAI Code of Practice (finalized in 2025, with chapters on transparency, copyright, and safety-and-security) as a voluntary route to demonstrate compliance. Signing it earns you what the AI Office calls increased trust — enforcement focused on your adherence to the Code rather than open-ended audits, and good-faith signatories aren’t going to get hit the instant the clock strikes if they’re visibly implementing.

Two things not to misread. First, the Code is not legally binding and not a safe harbor — adherence doesn’t immunize you from fines, it just gets weighed in your favor when the AI Office sizes one. Second, if you don’t sign, you don’t escape the obligations; you just have to demonstrate compliance some other way and explain your method to the regulator, which is more work, not less. The Code is the paved road. You can go off-road, but you still have to arrive at the same destination and show your route.

Is There a US Equivalent? Same Answer as Last Time: No

If you read the liability post, this will sound familiar, because the divergence is the same shape. There is no US federal equivalent to the AI Act’s GPAI regime. The 2023-era federal push toward AI safety obligations was rolled back; the 2026 posture is deregulatory, leaning on voluntary commitments and existing sectoral authorities rather than a horizontal statute with a dedicated enforcer and turnover-based fines. Some US states are moving on their own, but there is nothing that hands a regulator the power to demand a frontier lab’s documentation, evaluate its model, and pull it from the market.

So the same Brussels-effect logic applies as with GDPR and the PLD: a global lab is not going to maintain one model for Europe and a laxer one everywhere else. It’s cheaper to build to the EU bar once. Which means the AI Act is quietly setting the global floor for how the most capable models are documented and secured — decided in Brussels, exported by economics, with no US vote required. Between this, the Product Liability Directive, and the Cyber Resilience Act’s secure-by-design duties, the EU has spent 2026 becoming the de facto standards body for software and AI safety, and the US has spent 2026 explicitly declining the job.

My Read: What Security and AI Teams Should Actually Do

Strip the compliance vocabulary and August 2 is an operational to-do list. Most of it is what a serious AI security program should already be doing; the difference is that “we’ll get to it” now has a regulator attached.

1. Know your role before the AI Office decides it for you. Provider, downstream provider, or deployer changes everything about what you owe. And watch the fine-tuning tripwire: modify a model enough and you become a provider. Map every model in your stack — including the open-weight one running on a box in the research team’s closet — to a role, on paper, now. This is the AI Strategy Vacuum made concrete: you cannot govern models you haven’t inventoried.

2. Treat red-teaming and evaluation as evidence, not theater. If you touch a systemic-risk model, adversarial testing is now a legal duty — so do the real version. External, adaptive, adversary-incentivized, documented. The report you generate is both your safety evidence and, someday, an exhibit. A red-team artifact that lists a capability you then shipped anyway is worse than none.

3. Weight security is compliance now — protect the crown jewels like it. The Act names cybersecurity protection of the model and its infrastructure. Threat-model weight exfiltration explicitly: access controls, egress monitoring, insider risk, the supply chain around your training and serving infra. The model-is-the-attacker incident showed what happens at that layer with no telemetry. Instrument it.

4. Build the incident-reporting muscle before you need it. Serious-incident reporting to the AI Office is on staggered timelines. That means you need detection, classification, and a reporting runbook for model incidents — not just your classic IT SOC. If you can’t tell when your model did something reportable, you can’t report it in time.

5. Use your upstream provider’s new transparency duties. If you’re a deployer, your model vendor now owes you integration and risk information within days. That’s not just paperwork — it’s the raw material for your own security review of a model you didn’t train. Ask for it. In writing.

6. Sign the Code of Practice with clear eyes. For providers, it’s the lower-friction path and it counts in your favor. Just don’t mistake it for a shield, and don’t sign chapters you can’t actually implement — a broken commitment is worse than an honest “we’re doing it our own way.”

7. If you’re a US provider, don’t read the home-front deregulation as cover. Your government stepping back from AI rules shrinks your EU exposure by exactly zero: place a model on the EU market and the AI Office can still demand its documentation, evaluate it, fine you, and pull it. Same Brussels-effect logic as the liability piece — the EU bar is the one you end up building to, whatever Washington decides.

So What

For a decade, “we’re all in on AI” has been a slide, not a system. The AI Act is the EU deciding that if you place the most capable models on Earth into a market of 450 million people, you will document them, test them adversarially, secure their weights, report when they go wrong, and answer to someone with the power to fine you and unplug you. That’s not red tape. That’s the deal every other high-consequence engineering field accepted long ago, arriving — late, imperfect, but arriving — for AI.

I’ve spent this blog arguing that the model and agent layer is under-secured and barely instrumented. It is strange to watch a regulator show up and mandate a chunk of exactly that. I’ll take the win and keep my skepticism for the enforcement — because a rule is only as real as the first fine, and we’re about to find out how serious the AI Office is.

And remember this is only the first jaw. The regulator can fine you; the Product Liability Directive lets the person you harmed sue you, no fault required, EULA disclaimers void, in December. One summer, two doors, both now open. Build for both.

02-pincer-graph

Figure 2. The pincer — an insecure or non-compliant AI product reaches realized liability by two convergent tracks: the AI Act’s regulatory jaw (the AI Office, from August 2) and the Product Liability Directive’s civil jaw (strict liability, from December).

Stay paranoid. Red-team for real. Protect your weights.

Further Reading:

Questions or feedback? Reach out via:

Need help getting your models and AI systems ready for the AI Act era? VULNEX offers:

  • AI system and model security assessments (evaluation, adversarial testing / AI red teaming, open-model threat modeling)
  • GPAI role mapping and governance gap analysis (provider vs. deployer, fine-tuning exposure, model inventory)
  • AI incident-response readiness and model-layer telemetry design
  • Secure-by-design review for AI-driven products and agentic systems

For AI security strategy — where model and agent risk meets board-level decisions — see vulnex.ai.

Contact: info@vulnex.com

Posted in AI, Privacy, Security, Technology | Tagged , , , , , | Leave a comment