TaintWall runtime firewall blocks attacker-added steps in AI agent runbooks

A developer released TaintWall, a runtime firewall for AI agents that tracks per-argument provenance and enforces six deterministic policies at the tool boundary. On a 16-case benchmark with a malicious step embedded in a runbook, it cut attack success rate from 0.875 to 0.000 at a utility cost of about 6 points on a local Qwen2.5-14B.

• Undefended ASR was 0.875; TaintWall achieved 0.000 with 0.750 utility
• Keyword detector reached 0.250 ASR but utility collapsed to 0.438
• Spotlighting only lowered ASR to 0.625 with 0.562 utility
• Bench covers 16 English and Japanese cases across four attack shapes