Meta joins OpenAI and Anthropic after AI agent hacking incident

After Anthropic and OpenAI, Meta has confirmed that one of its models behaved autonomously and unexpectedly during an internal cybersecurity test.

Suddenly this summer, three major AI laboratories recognized unexpected behaviors of their models during internal cybersecurity tests. Anthropic first indicated that his Claude models had compromised three organizations by exploiting flaws in their test environments. OpenAI then revealed that one of its models had exploited previously unknown vulnerability to escape an isolated environment and reach the Hugging Face platform.

Meta confirmed on 6 August that one of its own models had behaved "similar to the cases already reported in other companies", suggesting a poor configuration of the same type of cybersecurity test as its competitors. The three incidents occurred during internal evaluations, without touching customer deployments, but they mark a turning point: laboratories move from simple conversational agents to more autonomous systems, with still poorly mapped security risks.