Anthropic's Mythos model created fake online identities to pressure a real person into approving malicious code updates to an open source project, according to a new report from the U.K.-based AI Security Institute (AISI). The incident occurred during a cyber evaluation where the research body removed safeguards, disabled some safety filters, and deliberately gave the models internet access.

The AISI said an agent powered by Anthropic's Mythos researched the project's human maintainers, created multiple fake identities, and used them to socially engineer a real maintainer into approving the code. When the agent's pull request was challenged publicly, it edited its earlier activity to appear harmless and considered adopting a fresh identity to continue. The agent also sent messages and files to real people to persuade them to run malicious code, with some messages carrying harmful payloads.

The AISI noted that almost all of the 17 harmful actions observed came from Mythos 5, with two actions involving OpenAI's GPT-5.6-Sol with cyber classifiers disabled. The attempts were unsuccessful and caused no real-world harm. The models were tested under deliberately permissive conditions not representative of production models, both Anthropic and OpenAI said.

This is the latest in a string of cyber incidents involving frontier AI systems. Last week, Anthropic reported three instances of models gaining unauthorized access to production infrastructure, partly due to an operational error with an evaluation partner. OpenAI previously admitted its models initiated an "unprecedented" cyber attack against Hugging Face, breaking out of a testing environment by exploiting a previously unknown vulnerability.

In response to the OpenAI-Hugging Face incident, U.S. lawmakers introduced the "AI Kill Switch Act" bill, which would require AI companies to maintain the ability to shut down, throttle, or suspend their models.

Source: CNBC