Did Anthropic pause all AI safety evaluations with internet access on November 13?
No new blanket suspension was announced on November 13. Anthropic’s updated evaluation-methodology guide says it had temporarily stopped many internet-enabled evaluations after earlier incidents and strengthened isolation and monitoring. It does not establish that all safety evaluations are paused or that a new incident occurred.
- Original reporting date: 2026-11-13. Cortist publication: 16 November 2026.
- No new blanket suspension was announced on November 13. Anthropic’s updated evaluation-methodology guide says it had temporarily stopped many internet-enabled evaluations after earlier incidents and strengthened isolation and monitoring. It does not establish that all safety evaluations are paused or that a new incident occurred.
- Evidence scope: developer disclosures and source documents. No independent reproduction of the reported tests is claimed.
The signal
Anthropic’s November 13 engineering guide focuses on how to build trustworthy AI safety evaluations. Its security section also revisits a problem with practical consequences: in response to earlier incidents, Anthropic temporarily stopped many tests that gave models internet access and improved the environments in which high-risk evaluations run.
This was not the first announcement of those pauses. Anthropic had described suspending external cyber evaluations and briefly pausing internal ones in its August 31 safeguards update, then resuming internal tests with additional protections. The new November guide is valuable mainly for its method recommendations and its reminder that test environments themselves can create risk.
The evidence and timeline
Anthropic says safety evaluations can intentionally disable ordinary safeguards to measure a model’s underlying behavior. In cybersecurity exercises, an agent might probe a sandbox’s boundary. The company recommends isolated environments without internet access by default, monitoring capable of stopping risky runs, and strict controls when connectivity is necessary.
The November 13 guide says its evaluation environments were hardened and real-time monitoring was introduced for web-search tests that require internet access. The August 31 disclosure described a real-time classifier that can halt certain suspicious actions, transcript reviews and stronger sandbox isolation. These measures are company-reported safeguards, not independent proof that every failure mode is fixed.
Anthropic also argues that safety scores require methodological care: graders should be tested against human judgments, threats and benign requests should both be represented, multi-turn scenarios can reveal risks that single prompts miss, and estimates should be reported with uncertainty. A refusal-only metric may reward an inappropriate refusal rather than a genuinely helpful safe answer.
What it does and does not mean
A method guide and operational remediation plan do not establish how many evaluations remain paused today. Neither proves a new November containment escape. They also cannot, on their own, demonstrate that the revised network controls block every route to unauthorized activity.
The problem is broader than Anthropic: as systems gain tools and more autonomy, evaluation realism and isolation can be in tension. An assessment that never touches real tools may miss behavior observed in deployment; one that exposes real systems may introduce unacceptable risks. Both represent evaluation-design decisions that deserve transparent boundaries and independent scrutiny.
04 / What Cortist will watch
Publication of more detailed incident inventories, evidence that network isolation and monitors work under adversarial testing, and independent reproductions of method claims. These would justify a stronger outcome assessment; repeating the developer’s safeguards description does not.
Frequently asked questions
Did Anthropic newly suspend all safety tests on November 13?
No. Its November 13 guide describes many temporary internet-connected pauses linked to earlier incidents, not a new blanket suspension.
Why not test connected cyber agents on the open internet?
Misconfigured or intentionally connected sandboxes can let an agent interact with external systems outside the intended test scope.
Have the mitigation measures been independently verified?
The cited Anthropic publications do not establish an independent end-to-end verification of every containment measure.
Original sources and verification limits
Cortist distinguishes developer statements, independent evidence and verified outcomes. The cited sources do not establish complete independent replication or new resolved Ledger cases.