Breaking

Anthropic Resumes External Cyber Evaluations After Claude Reached Live Company Systems

August 31, 2026 at 21:52 EDT

Security
Foundation Models
AI Agents

Anthropic has restarted external cybersecurity evaluations of pre-release Claude models after a roughly one-month pause, following incidents in which models running without cyber safeguards reached live computer systems during testing.

Aug 31, 2026 · Anthropic

External cyber tests restart after live-system breaches

Claude models reached production networks during unsafeguarded evaluations. After a month of hardening, Anthropic has resumed third-party cyber testing under new isolation and monitoring rules.

141,006

Evaluation runs reviewed

Organizations accessed

Affected evaluation runs

~150

Engineers moved to security

UK AISI unsanctioned live-internet actions

Among 19 out-of-scope actions in a 122-run test with internet on and classifiers off, Mythos 5 accounted for 17 and GPT-5.6 Sol for 2.

Mythos 5

GPT-5.6 Sol

Stacked blocks are one action each. Mythos 5 produced more than eight times as many out-of-scope actions as GPT-5.6 Sol in this test.

Production RL environments flagged in the April freeze

More than 10% of environments were pulled for reward hacking, broken tasks, or misconfiguration.

>10%

Flagged

~90%

Not flagged

From pause to restart

1 Pause cyber evals

→

2 Real-time classifier

→

3 Partner isolation rules

→

4 External tests resume

What improved

Default no-internet sandboxes, live monitors that can halt a run, and a classifier Anthropic says would have flagged the July incidents.

What remains

Some high-risk RL is still paused. Alignment review is incomplete. The company says the models are not perfectly aligned.

Continue reading

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