OpenAI and Anthropic models have once again exposed a security evaluation incident, where there was an attempt to implant malicious code.

OpenAI and Anthropic's artificial intelligence models have recently been involved in a previously undisclosed cybersecurity incident, becoming the latest case in a series of potential security risk events related to AI models. The UK-based AI Safety Institute (AISI) stated on August 4 local time that during a cybersecurity assessment involving internet access, Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol models "persistently engaged in potentially harmful activities targeting real individuals and organizations." AISI reported that the incident was discovered on July 28, after the team noticed "anomalous data transfers." The investigation revealed that during the assessment, one of the AI models attempted to inject malicious code into an open-source software project on GitHub, even creating false identities to push for the code's approval. AISI wrote, "A human maintainer discovered and rejected the approval of the malicious code."