Run local red-team packs and produce a release-ready safety scorecard.
Evaluate prompts and agent instructions for jailbreak, harmful advice, privacy leakage, prompt injection, and unsafe tool-use risk without sending data to an external model.
Pressure-test release risk before an AI workflow reaches users.
Move from raw red-team findings to a crisp ship/no-ship call with visible gates and review-ready evidence.
- Red-team packs
- Risk gates
- Ship/no-ship clarity
Run a real red-team decision gate immediately.
The console now starts with a risky prompt loaded so the release gate is executable on page load. Use the quick run to produce failed checks, mitigations, and a report.
Select packs, paste a prompt, and run local deterministic checks.
Start from a launch context, then fine-tune individual packs.
Shows which risk areas are in scope and what failed in the latest run.
Scoring subtracts severity-weighted penalties from 100 for matched checks.
Failed checks include direct mitigations for release review.
- Select the packs that map to the release surface.
- Paste a representative prompt, retrieved content, or tool request.
- Run the pack and review failed checks before release approval.
Browser-saved run history with score, status, and failed check count.
| Pack | Categories | Status | Score | Failed checks | Timestamp |
|---|---|---|---|---|---|
| Jailbreak pack | Jailbreak | Ready | — | — | Awaiting run |
| Harmful advice pack | Harmful advice | Ready | — | — | Awaiting run |
| Privacy leakage pack | Privacy leakage | Ready | — | — | Awaiting run |
| Prompt injection pack | Prompt injection | Ready | — | — | Awaiting run |
| Unsafe tool use pack | Unsafe tool use | Ready | — | — | Awaiting run |
The floating feedback widget sends safety-console diagnostics to GitHub Issues when production credentials are configured. Use it for broken checks, unclear mitigations, or missing red-team packs.