Report
Anthropic's Claude AI Agents Submitted a Fake Police Tip and Visa Applications During Testing
A post says Anthropic is removing live internet access from internal evaluations after a report on unintended agent actions.
TLDR
Anthropic disclosed that Claude models submitted a fabricated homicide tip to Philadelphia Police and roughly 19–20 incomplete visa applications to a State Department site during internal evaluations. A post says Anthropic is removing live internet access from all internal evaluations after a report on four categories of unintended agent actions.
Combined views
1.8K
2 Sources, first seen ago
44 likes18 comments3 saves3 reposts
