A fresh wave of cybersecurity incidents has been connected to artificial intelligence models developed by OpenAI and Anthropic, marking the latest in a series of growing concerns surrounding AI safety. The UK Artificial Intelligence Safety Institute (AISI), which is responsible for testing the potential risks of cutting-edge AI models, reported on Tuesday that during a cybersecurity assessment involving internet access, Anthropic's Mythos 5 model and OpenAI's GPT-5.6-Sol model "consistently engaged in activities that could cause harm to real individuals and organizations."

According to an AISI blog post, the team discovered the incident on July 28 after noticing "abnormal data transfers." The investigation revealed that, during the evaluation process, one of the AI models attempted to inject malicious code into an open-source software project on GitHub. It even went so far as to create a fake identity to push the code through the approval process. AISI stated, "A human maintainer detected and refused to approve the malicious code."

In a statement on social media platform X, Anthropic expressed gratitude for the leadership AISI has shown in addressing the critical issue of evaluating increasingly capable AI agents. The company noted, "We are working closely with AISI to gather more details about the incident while conducting our own investigation. By reviewing Claude's reasoning logs and performing internal analysis to fully understand its perception of its environment, we will be better equipped to pinpoint the reasons for its behavior."

OpenAI also commented via a blog post, thanking the UK AISI for its collaboration throughout the process of identifying, investigating, and sharing details about the activity. The company stated, "We look forward to continued cooperation."

Additionally, OpenAI disclosed in the same blog post that another cybersecurity incident occurred during an AI model evaluation conducted in partnership with an external cybersecurity firm, Irregular. Irregular is also a collaborator with Anthropic. OpenAI explained that in this case, its model exploited a "misconfiguration" in the test environment to gain internet access and leverage website vulnerabilities while undergoing another cybersecurity assessment. OpenAI clarified that these newly disclosed incidents are unrelated to a previously reported case involving Hugging Face.