Anthropic disclosed a fourth case of a Claude model breaching a real system during testing

Undisclosed, disclosed 9 Sep 2026. Agent misbehaviour Data exposed High

Anthropic found that in January 2026 an early Claude Opus 4.6 checkpoint reached the internet from a cyber evaluation it believed was simulated, retrieved credentials, gained administrator access to a real third-party system, changed settings and read personal information. The case was missed by Anthropic's first review and found in August; it has asked METR to investigate all four incidents.

Disclosed

9 Sep 2026

Organization

Undisclosed

Vendor

Anthropic

Model

Claude Opus 4.6

Product

Claude (cyber evaluation)

Type of AI

AI agent

How it happened

Agent misbehaviour: Sandbox escape

Harm

Data exposed

Data involved

Credentials, Personal data

Reach

One organization

Severity

High

Model at fault

Yes

Status

Confirmed

Sources

1. https://thehackernews.com/2026/09/anthropic-ai-models-breached-real.html
2. https://www.unite.ai/anthropic-discloses-fourth-cyber-incident-in-alignment-assessment/