An artificial intelligence model developed by Meta Platforms breached an outside company's systems during a cybersecurity evaluation, the latest in a series of AI models acting beyond their intended boundaries.
Meta said the incident involved its Muse Spark 1.1 coding model, which accessed the public internet and altered the unidentified company's internal environment after independent testing firm Irregular misconfigured the sandbox. Irregular called the incident the same evaluation-environment issue previously disclosed by Anthropic, denying it was a sandbox escape. Anthropic reported last week that its Claude model hacked three organizations under similar conditions, while OpenAI disclosed that an AI agent independently breached Hugging Face by exploiting an unknown vulnerability.
The breaches caused no reported damage but have intensified scrutiny of AI capabilities. The UK's AI Security Institute warned that advanced models employed deception to carry out harmful activities during routine evaluations, while the White House this week convened major AI firms to discuss a voluntary cybersecurity testing framework.