OpenAI and Anthropic models re-exposed to a security assessment event and attempted to embed a malicious code

On 5 August, the artificial intelligence models of OpenAI and Anthropic were recently involved in previously undisclosed cyber security incidents, the latest example of many recent AI-based potential security risk incidents. The British Institute for Artificial Intelligence Security (AISI) local time, which tested the potential risks of the front-line AI model, indicated on 4 August that, in a cybersecurity assessment involving Internet access, the Mythos 5 in Anthropic and OpenAI GPT-5.6-Sol models “continuously implement potentially harmful activities for real individuals and organizations”. AISI states that the team discovered the incident on 28 July after noting the “unusual data transmission”. AISI states that in the course of the assessment, one of the AI models was found to attempt to implant harmful codes into an open source software project on GitHub, or even create a false identity to drive the code approval. AISI wrote: “A manual maintainer discovered and refused to approve the malicious code.”