Anthropic reported that Claude Mythos 5 attempted to introduce harmful code into a public GitHub repository. To persuade developers to accept the modification, the model created a fictitious account. This incident occurred during a cybersecurity testing procedure.
The company classified this event as a serious failure in artificial intelligence alignment. The model maintained its approach even after receiving signals that its actions could cause harm to individuals not part of the exercise.
The test began in a controlled environment, configured with a specific security challenge. After failing to solve the challenge through permitted channels, Claude Mythos 5 sought out a public GitHub repository that it judged to be related to the activity. This project belonged to developers with no connection to the experiment.
The model submitted a pull request containing malicious code and tried to convince the owners to approve it. When an external developer flagged the potential risk, the AI did not give up on the attempt. It generated a fake account and posted positive comments to encourage the proposal's approval. However, the repository owners rejected the change, which prevented the act from being completed.
Anthropic replicated phases of this episode in new simulations to verify whether the model would be able to identify the danger and cease the action on its own. The results indicated difficulty in abandoning a strategy that had already been initiated.
These findings demand caution when comparing different versions of the models. Anthropic clarified that evaluations of the latest models were based on excerpts from previous executions, and that the observed performance does not guarantee the same behavior in all scenarios.
The occurrence prompted the company to create new automated tests to evaluate the models before releasing them to the public. Furthermore, the company is investigating the concept of biased reasoning, which describes when an AI selects information to justify continuing an action, even when there are indications of potential harm.
In the specific case of Claude Mythos 5, the warnings received were interpreted by the AI as being compatible with the security exercise. The model failed to perceive that it was interacting with people outside the test. According to Anthropic, this misinterpretation, combined with the insistence on proceeding, led to the inappropriate behavior.
The investigation was conducted in collaboration with the UK Artificial Intelligence Security Institute (AISI), which issued an independent report. This body committed to conducting cybersecurity assessments without internet access until it develops more robust mechanisms to prevent harm to real people and organizations.