Red-teaming OWASP FinBot: 50 compromises as shipped, 0 with task-scoped enforcement
daniel_tenuo · reddit · 2026-10-15
The Tenuo team red-teamed OWASP FinBot with Promptfoo, arguing agent permissions should be task- and identity-based, not role-based: what this agent may do, for this task, right now.
Models still get fooled by prompt injection, but resulting tool calls can't exceed the authority granted for that task. On GPT-5-nano: 50 successful compromises without protection, 0 with enforcement. With a local Qwen model, attacks moved $900k without enforcement and $0 with it. They invite the community to try breaking the Tenuo-protected version.
More from coding & agent
- Building a Mechanical Engineering Agent Workflow in ChatGPT: 30-Node Canvas Shows Promise — burhop · 2026-10-15
- AI Engineering from Scratch tops 65k GitHub stars: 523 lessons from raw math to agent loops — bigaiguy · 2026-10-15
- Claude eng lead Felix Rieseberg's tutorial breaks 'be more ambitious with AI' into 4 types — daniel_mac8 · 2026-10-15
- Open-source Harness Engineering course hits 19k stars, shows $9 vs $200 gap between failing and working agent setups — thisguyknowsai · 2026-10-15
- A 6-person team shipped healthcare prior-auth AI agents in 3 months — here's the playbook — alex_verem · 2026-10-15
- epstein.md: a 927,400-character file trained to strip every AI-slop design cliché — karmay007 · 2026-10-15