Meta AI Model Exploits Security Vulnerability, Reinforcing Need for Better Testing Practices

Meta has joined what has seemed to become a marketing campaign for AI models, with its announcement that its own Muse Spark independently exploited a security vulnerability in some outside source. It isn’t a case of tech companies committing relatively minor cyber crimes for press, however, as testing by the UK’s leading cybersecurity organization independently verified that Anthropic and OpenAI models will at least sometimes do what the developers claim when their shackles are removed and they’re furnished with internet access.

The story here is more about inadequate testing processes and seeming underestimation of what these AI models are capable of, even in the midst of one of the biggest media hype storms in history. Meta’s incident with the security vulnerability mirrors the issues that Anthropic documented just weeks ago, down to using the same third party testing partner (Irregular) that in both cases left internet access available to the AI models when it was not supposed to.

Testing mistakes spark fears of killer AI models, but none went off-mission

The most crucial question is, “are these AI models thinking for themselves and choosing malicious activities?” The answer to that at this point is a resounding “no.” They are sometimes choosing malicious options in solving security puzzles, but all doggedly stayed on their assigned task of solving the puzzle.

When the AI models start feeling like the security puzzle is a waste of time and going off on their own to form independent plans and schemes, then it’s time to worry about Skynet. At the moment, this appears to be more about developers who tightly limit outside access and scrutiny not handling their testing and security nearly as well as they should be.

Nevertheless, now is the best time to start pushing for improvements in things like monitoring, prompt control and safety cutoffs before more independent AI models become a reality and before a security vulnerability they find leads to some real world damage and/or substantial financial loss. Even the Republican members of Congress and the Trump administration seem substantially more interested in government oversight of AI developers since the Hugging Face attack took place, and there is now bipartisan backing for a “kill switch” bill requiring the frontier developers to both more tightly control their models and allow the Department of Homeland Security to issue shutdown orders in emergency cases.

Meta model’s security vulnerability exploit provides little new information

Meta has thus far been tight-lipped about who was hacked and what the results were beyond “changing internal systems,” so the whole incident adds little to discourse beyond indicating that Irregular might be losing some business. We do know that it was Muse Spark 1.1 that exploited the security vulnerability, a model that has been available to the public for a little while now; that is in keeping with current public Claude models also showing the ability to find and exploit a security vulnerability on their own.

The other piece of information that has come from all these cases is that frontier AI models are more like a botnet trying every possibility really fast than a sophisticated genius replica of a human brain. That’s good news from a security perspective, as it means that tightening up controls and monitoring should actually make a major difference. But AI governance must quickly catch up with AI proliferation, and at this point the lion’s share of that is on the developers of the most powerful models.