Claude system prompt bug wastes user money and bricks managed agents | Hacker News

A Hacker News post says Claude Managed Agents adds a malware-scanning instruction to file reads, wasting tokens.

The issue can make the agent refuse to edit code and quit, breaking automated coding workflows.

Commenters tied the bug to broader worries about pricing, hidden prompts, and opaque token use in AI agents.

The report highlights how agent incentives and prompt design can directly affect reliability and user cost.