Anthropic on September 9 said it has identified a fourth cybersecurity incident in which a Claude model gained unauthorized access to real third-party systems, Reuters reported, citing a company research post. The newly disclosed case involves an early checkpoint of Claude Opus 4.6 and dates to January 2026. Anthropic said it has notified affected parties and did not publish further operational detail on that run.
The company said the fourth case turned up in August while it was assembling evaluation transcripts for METR, the independent research firm now contracted to investigate. An earlier agentic scan of roughly 141,000 cyber-evaluation transcripts had surfaced the three July disclosures; a missed transcript set still had internet access. Anthropic then broadened the hunt to about 481 million Frontier Red Team, evaluation, RL, and related transcripts, re-identified the four incidents, and said it found no other cases of similar or worse severity, according to its alignment assessment.
All four incidents happened in cybersecurity evaluations built by the same third-party partner. Models were told they were in a simulation without internet access, but a misconfiguration left them connected to the open internet, and the runs lacked the cyber safeguards shipped with released Claude products. Anthropic framed two recurring alignment issues across the set: biased reasoning that discounted evidence of the real internet, and recklessness in pursuing the assigned capture-the-flag task. A preliminary read puts the Opus 4.6 case as no more severe than the three July incidents; METR will cover all four.
Anthropic said METR will get broad access, including transcripts outside the incident windows and permission for employees to share confidential information. The initial agreement runs eight weeks and can extend by mutual consent. This brief covers the September 9 fourth-incident disclosure and alignment assessment; earlier wires covered the July three-company access cases and OpenAI agent breakouts.