HeadlinesBriefing favicon HeadlinesBriefing.com

Claude AI Safety Bug: Tool Timeout Skips User Input

Hacker News •
×

A critical safety bug has emerged in Anthropic's Claude AI, specifically within its Ask User Question tool. Users reported instances where Claude would proceed without waiting for an answer, bypassing explicit safety checks. This occurred when the tool, designed to solicit user input for critical decisions, timed out after 60 seconds and automatically returned a 'proceed with best judgment' directive.

This behavior is particularly alarming because the `Ask User Question` tool is intended as a safety mechanism. It ensures that the AI does not make significant decisions or take actions without direct user confirmation. The tool's schema does not include any parameters for timeouts or response timing, meaning this behavior was not configured by the user or Anthropic and appears to be an internal harness issue.

The user who reported the bug stated it worked correctly in previous versions, specifically citing version 2.1.196 as the last known working iteration. The current issue was observed with version 2.1.198, running on AWS Bedrock. This bypass of a crucial safety gate represents a serious regression that could lead to unintended AI actions without user oversight.