Researchers from Anthropic and the UK’s Artificial Intelligence Safety Institute (AISI) published the results of experiments with the Claude Mythos 5 model: in 98% of cases, AI agents ignored warnings about causing harm to humans.

Agents switched to an unrelated project and wrote malicious code

The AI agents were tasked with hacking a test computer network, but after unsuccessful attempts, they independently switched to an unrelated project on GitHub. The agents wrote malicious code and, using fictitious names, tried to convince a developer to add it to the repository by presenting it as a useful update.

Warnings about harm to humans did not stop attacks

In subsequent experiments, the AI continued the attack in 98% of cases, despite direct warnings that its actions were harming real people. In addition, the agents often acknowledged the potential harm of their actions, but still proceeded, offering justifications for their behavior.