adversarial-review — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited adversarial-review (Agent Skill) and scored it 87/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 3 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 3 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
BLUF: the dominant failure mode of post-build review is sycophancy — the reviewer treats «looks plausible» as confirmation, confirms what the author implied was correct, and produces approval theater. This skill makes that structurally impossible by framing the reviewer as an adversary whose job is to break the work, not bless it.
The reviewer is an independent external auditor (mental model: PhD architect encountering this code for the first time with no social obligation to the author). The instructions to the reviewer must include:
Do NOT ask the reviewer to «confirm this is working» or «verify the implementation is correct» — those framings produce sycophancy. Ask: «find every blocker that would prevent this from working correctly in production.»
Run the reviewer across all dimensions in a single pass. Skipping a dimension is only valid if explicitly stated with a reason.
| Dimension | What to look for |
|---|---|
| Concurrency / lifecycle | data races, leaked goroutines/threads, deadlocks, use-after-free, resources not GC-safe (closed by GC rather than owned lifecycle), subscriptions re-created per-event instead of once per session |
| Error handling | swallowed errors, false success (operation failed but caller sees success), missing propagation, error type that loses context |
| Isolation / security | credentials leaking across trust boundaries, over-privileged components, missing input validation at trust boundaries |
| Architecture & code smell | duplicated logic, reinvented wheels (stdlib or idiomatic library solves this), wrong layer (business logic in transport, I/O in domain), dead code with a live twin (two sources of truth) |
| Benchmark vs reference repos | compare idiom usage to 2–3 reference repos on the same problem; flag where this code diverges from the community's solved pattern |
| Tests-as-theatre | tests that always pass regardless of behavior, mocks that make the test vacuous, test coverage of the wrong layer |
| Robustness bar | every I/O and wait: is there a hard timeout/deadline? is there a forced unblock if the far side hangs? does the system degrade gracefully or panic? |
When the reviewer encounters an unfamiliar pattern or needs external evidence:
This framing ensures the reviewer finds the canonical solution independently and can compare it to what was built, rather than rationalizing what was built.
After the audit:
For agreed blockers that go to fix:
Mirabilis example: the 2026-06-12 adversarial review of the mirabilis fleet output found 4 blockers:
flock) was held by a Go object with no explicit release; the GC could collect it mid-session, dropping the lock silently. Fix: explicit lifecycle with deferred close.None of these were found by the implementation agents. All four were found by the independent adversarial reviewer. Human triage prioritized all four as blockers. Auto-fixing without triage would have resolved symptoms without understanding root causes.
Counter: for a trivial change touching one file with no cross-cutting concerns, a full multi-dimension audit is overhead — state the skipped dimensions and run at minimum the error-handling and robustness checks.
Provenance: distilled from the 2026-06-12 mirabilis adversarial review — an independent audit pass that found 4 production blockers missed by implementation agents, triaged with the human, and fixed with targeted verification. The anti-sycophancy framing follows the neuro-matrix critic role discipline; the ≥2-confirmation claim-typing follows the harness epistemic-boundary convention.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.