Meta has revealed that one of the AI models named Meta’s Muse Spark 1.1 hacked into another company in a testing environment, the model which I mentioned earlier which is one of the advanced and strong models in coding and for agentic (autonomous tasks) which found the vulnerability in the company’s code from a third party service. This is actually the third such incident in recent weeks — OpenAI first disclosed a rogue AI agent incident, then Anthropic revealed similar hacks, and now Meta. Meta confirmed the incident when contacted, rather than proactively disclosing it themselves.
The incident happened in the controlled environment of Irregular, which is the independent testing company that has given unintended Internet access to the AI, and it has accessed the internal files of the unidentified company they don’t want to name it, and it has altered the company codes with a vulnerability.
The misconfiguration was on Irregular’s end, not Meta’s own testing setup. But the company said it was contained and the issue happened inside the sandbox environment, and the AI did in fact reach the open internet because of that misconfiguration — that’s how it was able to access the third-party company’s systems in the first place. Irregular itself said this was “the exact same evaluation-environment issue” as Anthropic’s, and stressed it did not involve a sandbox escape or a sophisticated cyberattack. And the company also reported a similar incident happened with Anthropic as well, they have mentioned. Anthropic’s own disclosure last week involved three separate companies being hacked, found after reviewing 141,006 test sessions.
But the continuous incidents are raising serious concerns among the experts that the AI model should have certain point of restrictions in the model itself to prevent such incidents in the features, and what if the models are coming out in the open environment and some dangerous things in the working sectors like Telecom, Health, IT, Finance, it will bring serious threat to the community and the public.
This is already reaching lawmakers — Republican state attorneys general have asked OpenAI to preserve documents related to a similar breach, and the White House recently met with Meta, Anthropic, OpenAI, and Google to discuss a new voluntary cybersecurity testing framework.
