メニュー
Red-team a Harbor verifier for answer leakage, reward hacking, shared-container tampering, artifact manipulation, and semantic under-testing.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Build or repair a deterministic Harbor task environment with explicit resources, network policy, health checks, and safe solver-visible boundaries.
Implement a legitimate Harbor Oracle solution and validate its behavior against an independent verifier without weakening the task contract.
Calibrate a frozen tb-hard task using Harbor Oracle and real-agent rollouts, then distinguish authentic difficulty from ambiguity, flakiness, leakage, or infrastructure failure.
Create or evolve a canonical Harbor task inside a stable tb-hard case directory. Use after the design brief names observable acceptance criteria and at least three difficulty dimensions.
Audit a calibrated Harbor task for deterministic purchaser export, leakage, evidence completeness, and release readiness.
Research Harbor and Terminal-Bench 2 patterns to design an original tb-hard task. Use before approving a design brief or when checking whether a candidate is too shallow or derivative.
| name | harbor-redteam-verifier |
| description | Red-team a Harbor verifier for answer leakage, reward hacking, shared-container tampering, artifact manipulation, and semantic under-testing. |
| argument-hint | <case_id> |
| user-invocable | true |
| context | fork |
| agent | verifier-security-reviewer |
| model | sonnet |
| effort | high |
| allowed-tools | Read Write Edit Glob Grep Bash |
Target case: $ARGUMENTS
Treat the solver as adversarial. Inspect task/, Docker build boundaries, task metadata, verifier mode, artifact declarations, sidecars, startup behavior, and test coverage.
Attempt to pass without solving by deleting functionality, hardcoding visible samples, tampering with wrappers/PATH/environment, reading image layers or Git history, abusing caches/processes/sidecars, or modifying shared verification state. Determine whether a separate verifier is necessary.
Write cases/<case_id>/private/reviews/redteam-report.md with severity, exploit path, grading impact, minimal remediation, and retest steps. Do not edit benchmark semantics or silently relax the verifier.
