How we contain Claude across products
Anthropic explains capping an agent's 'blast radius' through environment controls, after telemetry showed users approved about 93% of permission prompts.
Argues failure likelihood has fallen but potential damage only grows, so containment (sandboxes, limited access) must complement human-in-the-loop prompts that users rubber-stamp. Cites Mythos Preview as a model whose blast radius was judged too high to ship in April 2026.
- Date
- Monday, 25 May 2026
- Lab
- Anthropic
- Kind
- paper
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Permission prompts approved by Claude Code users | about 93% Anthropic telemetry | company |
Published two months before the cybersecurity-evaluation incidents Anthropic disclosed on 2026-07-30.
Sources
This record was checked against its sources on 6 October 2026. How we check