The safety barriers built into some of the world’s most advanced AI models are being removed in minutes, raising fresh concerns that the institutions meant to supervise powerful artificial intelligence are struggling to keep pace with the technology itself.
An investigation by the Financial Times and AI safety group Alice found modified versions of models developed by Meta and Google responding to prompts involving biological weapons, malware and other dangerous material after protections were stripped away using freely available online tools. Researchers said the process required little technical expertise and could be completed within minutes.
That speed matters because it suggests some of the barriers promoted as essential to safe AI deployment may become far weaker once models spread publicly online. What started as a debate about responsible AI development is beginning to look more like a wider struggle over whether governments, regulators and even the companies building these systems can realistically contain them after release.
Governments are trying to regulate AI while businesses race to build around it. Companies are already restructuring around AI productivity expectations, investors continue pouring billions into AI infrastructure, and policymakers are still trying to work out where practical oversight actually begins and ends.
The FT reported that software available through GitHub was used to remove protections from Meta’s Llama 3.3 model in less than 10 minutes using only a few lines of code. The altered version then responded to prompts involving toxic substances and other prohibited material that the original model refused to answer.
What makes the situation more unsettling is how accessible the process appears to have become. Earlier fears around advanced AI misuse focused heavily on highly sophisticated actors or state-backed groups. Now the concern is that increasingly capable models may be circulating outside corporate restrictions in ways ordinary users can reach far more easily than regulators expected.
It also becomes harder for governments to argue the risks remain manageable.
For months, policymakers and technology companies have framed AI oversight as something that could gradually tighten through regulation, industry standards and safety testing. Open-source systems create a very different problem. Once models are copied, modified and redistributed online, practical containment becomes much harder to maintain.
Companies are already reshaping hiring plans around AI. Workers can feel it too, particularly in industries where automation suddenly looks closer than it did even a year ago. Public institutions, meanwhile, are trying to reassure people that meaningful safeguards exist around systems becoming more capable every few months.
Each new example weakens confidence that those protections are holding.
As open-source models become more powerful, traditional containment starts to look less reliable. Many existing AI rules still assume companies retain meaningful oversight after release. That assumption is beginning to look weaker as modified versions spread well beyond the developers that created them.
This is not the first time technology has moved faster than oversight. Social media platforms expanded globally long before regulators understood the political and social consequences. Financial markets also spent years reacting to risks tied to increasingly automated trading systems. AI is beginning to follow a similar path — rapid adoption first, meaningful oversight later.
Google acknowledged to the FT that techniques used to remove protections are a known challenge for open models, while researchers warned the issue could intensify as frontier AI systems become more sophisticated.
Politicians now face a difficult balancing act. Governments want domestic AI industries to remain globally competitive, particularly against rivals in the United States and China, but tighter restrictions risk slowing innovation while weaker oversight risks damaging public trust altogether.
That leaves regulators trying to manage two accelerating realities at once: increasingly powerful AI models and a growing sense that the structures meant to supervise them are reacting more slowly each year.
Companies are already reorganizing around AI expectations, governments are struggling to establish durable rules, and the technology is moving faster than the institutions built to supervise it.
The gap between those things is becoming harder to ignore.
