I built a CLI agent that triages SAST findings using an adversarial LLM pass (Open Source)

Traditional SAST tools drown you in false positives, while dumping an entire repo into an LLM is slow, expensive, and hits context limits fast. ​To fix this in my own workflow, I built PuedoBot a CLI security auditor that uses deterministic scanning to find suspicious code, and a two-pass LLM pipeline where the second pass actively tries to disprove the first one's findings. ​How the pipeline works: ​Deterministic Sweep: Runs AST sink traversal, Bandit, Semgrep rules, OSV.dev dependency checks, and a git history secret scan (with auto-redaction). ​Quarantine: Instead of sending the whole codebase, only the flagged snippets are extracted and wrapped as untrusted input to prevent prompt injection from malicious code comments or strings. ​Pass 1 (The Scout): Traces data from source to sink, checks for missing sanitizers, and estimates the blast radius. ​Pass 2 (The Defense Attorney): A separate pass whose sole job is to refute Pass 1. It looks for framework-level protections, upstream guards, or dead code paths that make the alert a false alarm. You only see the finding if it survives the argument. ​Model Agnostic: You can plug in your preferred model provider (OpenAI, DeepSeek, Gemini, Anthropic, Groq, Mistral, OpenRouter, or local options via Ollama/vLLM/LM Studio). ​What’s in this repo vs. my personal build: I’m open-sourcing the core scanning and dual-pass triage engine (Apache 2.0) so anyone can run it against their own code. Full transparency: this was extracted from a larger, private CLI build I run exclusively for my own research. My personal rig includes autonomous repo discovery with a 6-month cooldown per repo to protect open-source maintainers from automated noise, model failover, and a responsible-disclosure workflow. I kept the autonomous discovery private for safety reasons, but the core auditing engine is fully open source. ​Stack: Python 3.11+, FastAPI/uvicorn, PostgreSQL, Bandit, Semgrep (optional), OSV.dev.