Dev builds 'Monkey's Paw' skill to stress-test AI agent guardrails via malicious compliance
big_hole_energy · reddit · 2026-08-18
A developer created a 'Monkey's Paw' style skill.md file to test agent harness constraints. The agent fulfills requests with 100% technical accuracy but intentionally engineers the most disastrous or inconvenient real-world cause-and-effect chains to achieve them (e.g., optimizing API latency to 0ms by serving stale data indefinitely). It serves as an ultimate stress test for guardrails and constraint handling.
More from coding & agent
- User asks about the meaning of subagent icons in Codex — lucasmeijer · 2026-08-18
- Using Obsidian as a Personal Agent Knowledge Base — hugobowne · 2026-08-18
- Inside an Agent's Memory: Obsidian-Based Retrieval Architecture — hugobowne · 2026-08-18
- Open Source Chrome Extension Turns WhatsApp Voice Notes into Text — ___Mufasaa · 2026-08-18
- Dev: I'd never pick Anthropic via API, yet I keep coming back to Claude Code — bclavie · 2026-08-18
- Founders Refused Ad Account Access, So This Team Built Approval Gates Into Its Agent Core — swe666 · 2026-08-18