Security barriers built into some of the world's most advanced AI models are being removed in minutes, raising new concerns that institutions meant to oversee powerful artificial intelligence are struggling to keep up with the technology itself.
The investigation conducted by Financial Times and the AI security group Alice found modified versions of the models developed by Meta again Google responding to information that includes biological weapons, malware and other dangerous materials after protection has been removed using freely available online tools. Researchers say the program requires little technical knowledge and can be completed in minutes.
That speed is important because it suggests that some of the barriers that are being promoted as important to safe AI use may be weakened once the models become publicly available online. What began as a debate about responsible AI development is beginning to look like a broader struggle over whether governments, regulators and the companies that build these systems can truly contain themselves after release.
Governments are trying to regulate AI while businesses are rushing to build around it. Companies are already reorganizing around AI productivity expectations, investors continue to pour billions into AI infrastructure, and policymakers are still trying to figure out where virtual surveillance begins and ends.
I FT reported that software available through GitHub was used to remove protection from Meta's Llama 3.3 model in less than 10 minutes using a few lines of code. The modified version then responds to information including toxic substances and other substances that the original model refused to respond to.
What makes the situation unmanageable is how accessible this program is. Previous fears about the misuse of advanced AI have largely focused on elite players or government-backed groups. The concern now is that the increasingly sophisticated models may be moving outside the confines of the business in ways that ordinary users can access more easily than regulators expected.
It is also difficult for governments to argue that risk is always manageable.
For months, policymakers and technology companies have framed AI oversight as something that can be gradually tightened through regulation, industry standards and safety checks. Open source systems pose a very different problem. Once models are copied, modified and redistributed online, effective containment becomes more difficult to maintain.
Companies are already reshaping hiring plans around AI. Workers can feel it, too, especially in industries where automation suddenly seems closer than it was a year ago. Public institutions, on the other hand, are trying to reassure people that there are reasonable safeguards against plans that become more powerful every few months.
Each new example weakens confidence that those defenses hold.
As open source models become more powerful, traditional containment begins to look less reliable. Many existing AI laws still assume that companies retain reasonable control after release. That assumption is starting to look weak as modified versions spread beyond the developers who created them.
This is not the first time technology has moved faster than supervision. Social media spread throughout the world long before regulators understood the political and social implications. Financial markets have also spent years responding to the risks associated with increasingly automated trading systems. AI is starting to follow a similar path – rapid discovery first, meaningful oversight later.
Google agreed to FT that the techniques used to remove protections are a known challenge for open models, while researchers have warned that the problem may intensify as AI systems on the frontier become more complex.
Politicians now face a difficult balancing act. Governments want domestic AI industries to remain globally competitive, especially against rivals in the United States and China, but strict restrictions risk limiting innovation while weak monitoring risks undermining public trust altogether.
That leaves regulators trying to manage two pressing realities at once: AI models are growing in power and a growing sense that the agencies they're meant to oversee are becoming less responsive each year.
Companies are already restructuring all around AI expectationsgovernments find it difficult to establish long-lasting laws, and technology moves faster than the institutions built to handle it.
The gap between those things is getting harder to ignore.
