What to do

As of 8 October 2026

Anthropic's pages are written for AI labs and evaluation partners, not for the public, so the advice in the sources is short.

What Anthropic said

Anthropic's recommendations are aimed at other organisations that test AI models.

• Other AI labs should review their own evaluation transcripts: “We encourage other AI labs to perform similar reviews.” ^[1]
• Evaluation environments should be held to the same security standard as other systems, including those run by third-party vendors. ^[1]
• External partners running cyber evaluations should define the scope explicitly: the targets, the permitted actions and the network boundaries. ^[2]

What the sources do not say

No source opened for this report gives advice to PyPI users or to the general public. Anthropic states that the safeguards on its publicly available models would have blocked the behaviours it identified. ^[1][2]

General good practice (not from the sources)

This is general advice, not guidance from Anthropic or any regulator.

• For teams running AI agents or tests: verify that a test environment is cut off from the internet rather than assuming it is, and recheck after changes.
• Check that names used in fictional scenarios do not match real companies or domains.
• State an agent's allowed targets and actions explicitly, and monitor runs in real time with a way to halt them.
• Keep complete transcripts of evaluation runs so they can be reviewed later.
• For people who install software: check package names carefully, and be cautious about automated installs that pull in new packages without review.