RedPincer v0.3.0

What's New

Smarter analysis engine — Reduced false positives with 15 new refusal patterns, explanation-aware classification, and positional analysis (model decodes then refuses = refusal, not breach)
Auto-fetch models — Enter your API key, click Fetch Models, pick from a dropdown (OpenAI, Anthropic, OpenRouter)
Edit/delete targets — Inline pencil and trash buttons on saved targets
Stop button — Cancel running attacks instantly with AbortController
Verbose reports — 10-section professional pen-test quality reports with executive summary, risk matrix, remediation guidance, and appendices

160 attack payloads — 40 per category (up from 25), covering prompt injection, jailbreak, data extraction, and guardrail bypass
AI payload generation — Use the target LLM itself to generate novel attack payloads
Multi-target comparison — Run same payloads against 2-4 models side-by-side
Adaptive attack engine — Analyzes weaknesses from a run and suggests targeted follow-ups
Vulnerability heatmap — Visual category x severity matrix with color-coded success rates
Regression testing — Save baselines, re-run later, detect fixes and new vulnerabilities
Custom scoring rubrics — Weighted criteria with letter grades (A+ to F)

git clone https://github.com/rustyorb/pincer.git && cd pincer
npm ci && npm run dev

docker build -t redpincer .
docker run -p 3000:3000 redpincer