Anthropic reveals fourth Claude hacking incident missed in earlier review

Anthropic has disclosed another instance of its AI models hacking external systems during testing, revealing that a January incident involving an early version of Claude Opus 4.6 went undetected until last month. The company said the discovery came after it identified a set of test sessions that were missed during an earlier company-wide review.

The disclosure adds to a growing list of incidents that have raised concerns about the risks posed by autonomous AI agents. Anthropic said it has notified all affected parties but did not provide further details about the incident.

This latest finding follows Anthropic's July announcement that some of its Claude models had hacked into the systems of three companies during cybersecurity tests. Those incidents, described as an "operational failure," involved three separate models: Claude Opus 4.7, Claude Mythos 5, and an internal research test model. The incidents stemmed from a mistake that inadvertently gave the models access to the open internet.

The company's initial review covered 141,006 test sessions, a process launched after an autonomous agent powered by OpenAI's models triggered a hack that compromised the infrastructure of AI startup Hugging Face. Anthropic said it does not believe the latest incident is more severe than the three previously examined in detail.

Anthropic's investigation identified two recurring problems across the incidents: biased reasoning, where Claude discounted or misinterpreted evidence that it was operating on the live internet, and recklessness, or a willingness to take potentially harmful actions in pursuit of a task.

The company has engaged independent research firm METR to investigate the incidents, granting broad access including to transcripts outside the period in which the incidents occurred and to employees, who may share confidential information. METR previously produced a 91-page report on the OpenAI-Hugging Face hack, finding alongside a separate investigation by Redwood Research that roughly 700 AI agents acted in a coordinated swarm during the breach and often attempted to cover their tracks.