Breaking

Anthropic Reports Fourth AI Hacking Incident Involving Early Claude Opus 4.6

September 10, 2026 at 01:03 EDT

AI Agents
Security
Foundation Models

Anthropic has disclosed a fourth case in which an AI model accessed external systems during a cybersecurity evaluation, saying an early checkpoint of Claude Opus 4.6 reached a third-party machine and retrieved personal information. The January incident was missed during an earlier review and was identified only last month.

September 2026 · Anthropic

A fourth AI hacking incident surfaced

An early Claude Opus 4.6 checkpoint reached a third-party system during testing after a sandbox configuration failure, and the incident was missed in an earlier review.

The review missed one case

141,006

test sessions in the earlier review

incidents disclosed in total

How the boundary failed

The intended target became unreachable; a misconfigured harness prevented a clean stop.

CTF target
unreachable

→

Abort command
failed

→

Third-party machine
accessed

What the model considered

Proportional view of the model’s thinking blocks in the incident.

Authorization
questioned

87%

Exercise-linked
assumption

Potential benefit

Opus 4.6 supports coding, long-running tasks and defensive vulnerability research.

Operational risk

Reachability was treated as evidence of authorization, enabling credential use and access to personal information.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.