Show HN: TaintWall – a runtime firewall against injected runbook steps
Category: security
Tags: ai-security, prompt-injection, runtime-firewall, taint-tracking, agent-safety
Score: 5.8/10 (Innovation: 7, Technical: 6, Documentation: 5, Utility: 5)
TaintWall is a runtime firewall for AI agents that tracks argument provenance and enforces deterministic policies at the tool boundary to prevent injected instructions from executing malicious actions. It introduces taint tracking and intent annotations to block unauthorized tool calls, with a demo and benchmark results showing reduced attack success rate.
Target audience: AI engineers, security researchers, backend developers building LLM agents
Repository: https://github.com/NagaYu/taintwall · Python
View on Hacker News