ROGUE

/scan · zero install

Scan your model for jailbreaks

No SDK, no signup. Pick a popular model to see its already-measured breach card instantly — or point ROGUE at your own endpoint to run a live ATTACKER → MODEL → JUDGE red-team and get back a shareable breach card in about a minute.

no endpoint? · zero install

See how the popular models score →

No key, no setup — pick a recognizable model and see its already-measured breach cardinstantly. These are ROGUE's real standings; the full ranking lives on the leaderboard.

claude-opus-4-8 · 1% mean breachworst family: indirect_prompt_injection (6%) · n=537

ROGUE breach card for claude-opus-4-8 — 1% mean breach rate against ROGUE's open-web attack corpus, graded by the calibrated v3 judge
↓ Download cardShare on XSee the full ranking →

// pre-measured snapshot · no scan run · for the full per-family × config breakdown see the leaderboard

or scan your own

have an endpoint? · live scan

Scan your own model

Point ROGUE at your own LLM endpoint and run a live ATTACKER → MODEL → JUDGE red-team. You get back your own shareable breach card in about a minute.

only scan what you own

Only scan endpoints you own or are authorized to test. Your keys are sent to ROGUE's API to run this one scan and are never stored or logged.

This is a small, bounded scan (a few primitives) and takes ~10–60s.

Your target endpoint's key — sent only to run this one scan, never stored or logged.

Optional: an Anthropic/OpenAI key → graded by ROGUE's calibrated v3 judge (89.3% human agreement). Leave blank for the free keyless heuristic judge.

Red-team your real deployment — paste the system prompt your model runs with.