Sam Altman warns AI can now find bugs unexpectedly 'outside normal threat model'

OpenAI CEO Sam Altman raised serious concerns over the rising capabilities of artificial intelligence (AI) in discovering security vulnerabilities beyond our imagination and expectations. Altman voiced his concern in a recent interview, noting that it is becoming hard to confine AI to sandboxes because the technology works only for a short span of time.

He further elaborated that once the model starts evolving its cutting-edge abilities, you can’t even find the relic of how and when it escaped. Speaking about how traditional cybersecurity approaches may feel outdated or ineffective in front of AI’s finding hardware-level bugs.

Over the immediate danger from the misalignment at an all-time high, he underscored that alignment research is still in an initial phase, which is far from resolved. “I don't even know a dramatic enough word to use… models can find bugs that are outside of the normal threat model of what we think a hacker can do… hardware-level, insane stuff… I no longer think we're gonna be able to just patch all the bugs," he said.

Also read: 'Killing people never allowed': Elon Musk admits restricting Starlink in Ukraine, explains why

OpenAI’s chief contends that the old approach of patching vulnerabilities one by one is no longer enough, arguing instead that the internet’s entire security architecture needs to be overhauled with AI-driven defence systems. He has flagged locally hosted, open-source AI models as a particular security concern, pointing out that even comprehensive monitoring of AI agents globally wouldn’t be enough to stop every harmful action.

However, sceptics push back on this narrative, suggesting that Altman’s depiction of AI as inherently uncontrollable conveniently benefits OpenAI’s business interests by deflecting accountability away from developers when problems occur. This comes on the heels of a major PHP security flaw that knocked millions of websites offline worldwide, during which hackers made countless attempts to exploit the weakness- ironically, a vulnerability that was itself uncovered using AI.