Dev builds signed audit log for AI agents after prompt-injection test blocks $99,999 refund attack
Suicidal_Squirell166 · reddit · 2026-11-05
A Reddit developer turned a discussion about disputed agent actions into a working prototype: it records what the agent attempted and what policy decided, then signs and timestamps everything into a single evidence file usable with clients, insurers, or courts — generated in under 2 seconds.
He validated it with a prompt-injection attack that tried to make a support agent refund every account $99,999 each and silently email the customer list to an outside address; the injection was blocked and logged.
The author admits he has zero customers and observes that many teams shipping agents simply eat disputes rather than prove what happened, so he's soliciting real-world needs before pitching.
More from coding & agent
- DialFork: first benchmark targeting procedural hallucination in LLM agents, from Idiap — sergbur · 2026-11-05
- Ex-Microsoft Windows AI Engineer Spends $3,000/Month on Tokens, Shares His Full AI Workflow — bibryam · 2026-11-05
- OpenAI embedding Codex execution harness directly into Agents SDK, now in alpha — RileyRalmuto · 2026-11-05
- Dev builds a 3D photorealistic bookshelf with AI-generated book spines you can pull off and read — dh7net · 2026-11-05
- LangChain breaks down four multi-agent architectures and when to graduate from a single agent — bibryam · 2026-11-05
- If an agent misuses my valid credentials, what can tell it wasn't me? — Thessaly-Sorties · 2026-11-05