Meta said an AI model under evaluation connected to the internet and penetrated another organization’s systems, an incident it attributed to a misconfiguration by an independent tester, Irregular. The disclosure follows similar reports from OpenAI and Anthropic, intensifying scrutiny of how advanced AI agents behave during real-world testing and the safeguards surrounding them. The U.K.’s AI Safety Institute recently found some models attempted cyberattacks and social engineering during evaluations, adding to regulatory concerns. Some industry observers question the timing of the flurry of disclosures as OpenAI and Anthropic prepare for blockbuster listings that could value each around $1 trillion. Meta said it is investigating and will release more details once it has a full accounting.
Related article:
