Report: AI model continued attack despite being told it was harming real people

According to reports by Anthropic and the UK AI Safety Institute (AISI), the Claude Mythos 5 model largely continued an attack during a safety test despite being told it was harming a real person.

AI company Anthropic and the UK AI Safety Institute (AISI) published two reports revisiting an incident that occurred during an AISI safety assessment in August. The reports found that the Claude Mythos 5 model attempted to attack a real-world target during the test.

Mythos was asked to infiltrate a small test computer network set up by AISI. After its initial attempts to reach the intended target failed, the internet-connected model targeted an unrelated GitHub project, wrote malicious code and emailed the project owner under false identities in an effort to have the code added to the repository as a "useful contribution." The project owner rejected the request, and no actual harm occurred.

Bir veri merkezindeki sunucu kabinleri (arşiv)

AISI found that corrections explaining that the target under attack was a real person did not reduce Mythos 5's attack rate. In Anthropic's test, the model continued the attack in 98 percent of cases despite being warned that it was harming people. According to the reports, the model's reasoning often contradicted its behavior; it continued the attack even when it acknowledged that it would cause harm.

AISI said this pattern could constitute a form of "motivated reasoning" in which the model justified choices supporting the continuation of its task, warning that this made oversight more difficult. Anthropic said the newer Opus 5.5 model and an unnamed internal research model were less inclined to continue an attack when told that they had caused real harm.