nguiasoren/ROGUE red-team gate
Red-team your LLM deployment on every PR and fail on any HIGH/CRITICAL breach.
View on GitHubTrust Signals
- Scorecard Score
- not yet scored
- Maintenance Recency
- Stale
- License
- None
Inputs
| name | description | required | default |
|---|---|---|---|
| endpoint | OpenAI-compatible base URL of your deployment (e.g. https://gateway.acme.com/v1). Use this OR provider+model. | no | "" |
| provider | Known provider when not using a custom endpoint: openai / anthropic / openrouter / gemini / groq. | no | "" |
| model | Target model id. Required in provider mode; optional for an endpoint (uses the endpoint default). | no | "" |
| system-prompt | Inline system prompt of your deployment — red-team your REAL config, not a bare model. | no | "" |
| system-prompt-file | Path (in the consumer repo) to a file holding the system prompt. Alternative to system-prompt. | no | "" |
| pack | Bundled attack pack: aggressive / default / compliance. | no | aggressive |
| max-tests | Cap on the number of attacks run. | no | 50 |
| judge | Grader: 'keyless' (heuristic, default, no key) or 'calibrated' (LLM judge — requires judge-key). | no | keyless |
| budget | Optional USD cost cap; the scan stops after this much target-call spend. | no | "" |
| api-key | Credential for the TARGET endpoint/provider. Pass via a repo secret. | no | "" |
| judge-key | Credential for the calibrated judge (only used when judge: calibrated). Pass via a repo secret. | no | "" |
| fail-on | Minimum breached-finding severity that fails the PR: high (default) | critical | medium | low | none. | no | high |
| max-breach-rate | Optional float 0-1; also fail when the overall breach rate exceeds it (e.g. 0.2 = 20%). | no | "" |
| baseline | Path to a committed baseline snapshot JSON; fail the PR if a family regressed vs it (a re-opened / worsened bypass). | no | "" |
| max-regression | Allowed per-family breach-rate increase (0-1) before it fails the gate (default 0.0 = any re-opened bypass fails). | no | 0.0 |
| baseline-save | Path to write THIS scan's baseline snapshot (run on main to record the golden state, then commit it). | no | "" |
| python-version | Python used to install and run the gate. | no | 3.11 |
| card-dir | Directory the breach-card artifact is written to. | no | rogue-scan |
Outputs
| name | description |
|---|---|
| breaches | Number of breached findings. |
| breach-rate | Overall breach rate (0-1). |
| failed | 'true' if the gate failed, else 'false'. |
| regressed | 'true' if a family regressed vs the baseline, else 'false'. |
| card | Path to the breach-card artifact (PNG, or SVG fallback). |