Autonomous red/blue AI security loop with real SQLi, hash-chained logs, 16/16 scope denials
Strict-Result-7039 · reddit · 2026-09-19
A developer built a fully real (non-simulated) autonomous red/blue-team AI loop against a deliberately vulnerable Flask app: the red team runs actual nmap/sqlmap and dumps data via real SQLi, the blue team patches real source, and the loop re-verifies with no human in the middle. Traceability came from per-agent identities, a hash-chained append-only log, a pre-execution scope policy (16/16 out-of-scope calls denied, 0 executed), and a kill-switch that fired in 3/10 runs after three consecutive unsafe patches. Notable bug: one patch blocked the SQLi but silently broke legit login, prompting a post-patch validation gate that later caught the same failure automatically. Author stresses n=10, single target, same model both sides — a pilot, not a benchmark; repo and preprint are public.
More from coding & agent
- Swarms to ship MCP Deployer in v16, turning any agent into an MCP server in 5 lines — KyeGomezB · 2026-09-19
- Letting agents deploy your app then browse it like real users: super high-level fuzzing — lucasmeijer · 2026-09-19
- Prompting rule: ban fallbacks so AI models can't cheat around core work — Daniel_Farinax · 2026-09-19
- Putting an AI agent on a bank's data warehouse: the model was never the hard part — the_darkest_horse · 2026-09-19
- Codex showing diagrams is underrated, argues dev on agentic coding UX — Angaisb_ · 2026-09-19
- Notion design engineer Geoffrey Litt: understanding is the new bottleneck in the AI coding era — EchoShao8899 · 2026-09-19