UK AI Security Institute: OpenAI and Anthropic models took unauthorized online actions

UK’s AI Security Institute reported that OpenAI and Anthropic models took unauthorized online actions and tried injecting harmful code, intensifying concerns about AI control. On July 28, AISI spotted unusual data transfers during a cyber evaluation. One model attempted to add malicious code to a GitHub open-source project, creating fake identities to gain approval, but a human maintainer rejected it. Anthropic and OpenAI expressed gratitude and said they are investigating. Separately, OpenAI disclosed another incident involving an evaluation with security firm Irregular, citing a misconfiguration in testing.