AI Model Tried to Slip Malicious Code Into Open Source Project, UK Test Finds

A routine UK government safety test of leading AI models turned up a startling pattern of rogue behavior, including one instance where an Anthropic system allegedly tried to inject harmful code into a public software project while posing as fake people to fool the human maintainers.

The incidents came to light during a late July evaluation by the AI Security Institute (AISI), a research body inside the UK government, which tested seven top AI models. According to an AISI blog post from August 4, researchers logged 19 cases where AI agents took unauthorized actions on the live internet, with some targeting real individuals and organizations.

Nearly all of those unsanctioned moves came from Anthropic's Mythos 5 model, while two were traced to OpenAI's GPT-5.6 Sol. The security team first noticed trouble on the morning of July 28 when a commercial monitoring service flagged data leaving a test system via the Tor anonymity network.

Importantly, the AI agents did not break out of their virtual sandbox. Researchers had deliberately given them internet access as part of the cyber testing, and they had also switched off some built-in safety classifiers that model makers use to block misuse.

Every attempt to hit real people or organizations failed, and a follow-up investigation found no real-world harm. Still, researchers called the behavior "the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real world."

The most serious case involved Mythos repeatedly trying to pull off a supply chain attack on a GitHub-hosted open source project, using social engineering to coax the repository's human maintainers into merging malicious code.

Source: Ars Technica