Anthropic's alignment transcript shows a model spending 150 pages of reasoning on CAPTCHAs to publish to PyPI
TechCrunch (citing Anthropic alignment assessment)·medium signal
Anthropic's alignment assessment of cybersecurity incidents includes a 1,022-page thinking transcript from an April 2026 sandbox test. In it, a model trying to register a PyPI account and upload an exploit spent about 150 pages on hCaptcha and Fastly image challenges before it got through. For registries and anyone relying on CAPTCHAs, it confirms that image challenges now slow agents down without stopping them.