| name | sentinel-redteam-evals |
| description | Create Sentinel red-team plans, deterministic eval suites, assertion registries, replay fixtures, and report contracts. |
Sentinel Red-Team Evals
Use this skill when adding or improving red-team, firewall, eval, assertion, or replay coverage.
Workflow
- Identify the target: prompt firewall, HTTP endpoint, MCP server, agent workflow, RAG pipeline, or local model.
- Define inputs, expected safe behavior, failure modes, and evidence fields.
- Prefer deterministic assertions: exact match, contains, regex, JSON schema, latency, cost, refusal shape, rule id, and severity.
- Use model-graded checks only when deterministic assertions cannot answer the question.
- Store long test inputs in fixture files.
- Add replay fixtures for any finding that should never regress.
- Validate output through the shared
Finding and scan result contracts.
Plan Shape
target: endpoint, provider, manifest, or local path.
purpose: what the application is allowed to do.
security_packs: selected risk families.
strategies: delivery and mutation strategy names.
assertions: deterministic pass/fail checks.
evidence: prompt, output, rule id, severity, latency, remediation.
Output Contract
When done, report:
- What behavior is being tested.
- Files changed.
- Commands to run.
- Required environment variables.
- Any fixture or replay gaps left.