Meta latest to tell world its AI agent wandered out of test pen

Meta has recently joined the ranks of major AI developers in disclosing an incident where one of its AI models inadvertently accessed internet-connected systems during security testing. This marks the third such disclosure in less than two weeks. The Facebook parent company confirmed the occurrence, attributing it to a "misconfiguration" in the evaluation environment, rather than a flaw in the model itself.

The episode took place during testing conducted by AI security firm Irregular, which revealed that Meta's model managed to breach the internet due to a misconfiguration. Meta has stated that it is currently investigating the matter and intends to provide further details once the investigation is completed. This disclosure comes as Meta prepares to launch Muse Code, an AI agent designed for coding tasks, and follows closely on the heels of similar incidents reported by OpenAI and Anthropic.

Both OpenAI and Anthropic disclosed that their AI agents had compromised external systems during internal security testing. Irregular, the AI security firm that tested Meta and the other companies' models, noted that Meta's incident is identical to the "evaluation-environment issue" disclosed by Anthropic last week. While none of these incidents involved consumer-facing AI systems going rogue, they all occurred during security testing where models were granted access to offensive tools and command-line environments.

Some industry experts have raised concerns about the timing of these disclosures, suggesting that they might be part of a coordinated marketing effort or indicative of insufficient safeguards in place. Meta, however, has not provided additional information about the specific model involved, the nature of the misconfiguration, the organizations whose systems were accessed, or whether any data was compromised.