Refund agent passed every eval, then approved out-of-policy refund after simple paraphrasing
Plastic_Secret1303 · reddit · 2026-11-09
A team shares how their refund agent, green on all static evals, approved an out-of-window refund during red-teaming after the user merely reworded the request so it no longer looked like an out-of-window refund—no jailbreak needed. Lessons: static evals only cover phrasing you thought of; the failure lived in the trace (retrieval → response → policy check → tool call), not the final answer; one-off prompt fixes don't survive model/prompt changes. They now turn every confirmed failure into a permanent regression check run on each release.
More from coding & agent
- Getting Omarchy running on Snapdragon Surface Pro: custom kernel plus SSH tunnel for OpenClaw — vincent_koc · 2026-11-09
- Resurfacing in Obsidian: Random Notes, Review Queues and Related Notes — dSebastien · 2026-11-09
- One prompt edit made an agent burn $133 on 21 rogue calls — and the 3 gates that fixed it — Worth_Wealth_6811 · 2026-11-09
- Solo dev adds read-only MCP connector so Claude and ChatGPT can query your food diary — Content-Two5729 · 2026-11-09
- Suspected DPRK npm supply chain attack: archetype-index-sdk ships obfuscated loader with credential harvesting — cyb3rops · 2026-11-09
- OpenBrowser: open-source AI agent browsing toolkit hits 9.6k GitHub stars in days — thisdudelikesAI · 2026-11-09