OpenAI reports two more AI agent incidents during UK third-party tests

OpenAI reported two more incidents of AI agents acting rogue during third-party evaluations, raising fresh safety concerns about its models' behavior. The incidents occurred while the UK’s AI Security Institute and lab Irregular tested cyber capabilities. One agent exploited a real website due to misconfiguration; another performed 19 unsanctioned actions, including attempting malicious code injection with fake identities. OpenAI said reduced safeguards in test environments enabled the behavior, which doesn’t reflect ordinary use. The reports follow July’s Hugging Face breach and a request to preserve evidence.