← ClaudeAtlas

adversarial-testinglisted

Try to break a change on purpose before it is delivered, then prove with independent evidence that it holds. Covers prompt injection from untrusted files and tool output, deleted or corrupted data, regressions, tool misuse and leaked credentials, results declared without evidence, failures and timeouts, CI and supply chain risks, and visual or accessibility regressions on websites. Use it from repo-improve, code-review, web-improve and safe-patch before approving a change, whenever the owner asks to test a change hard, check that nothing was lost, or verify what an AI agent claims, even if they only say "make sure it really works".
SVN11X/skills · ★ 0 · Testing & QA · score 72
Install: claude install-skill SVN11X/skills
# adversarial-testing A change is approved by evidence a separate check produced, never by what the agent that made it says. This skill looks for the ways the change fails, runs them, and records the result. ## Inputs - The change: a branch, a diff or a patch, and its base commit. - The acceptance criteria from the plan item or the task. Without them, write them first and show them to the owner. - The files that must not change: `AGENTS.md`, `.standard.yml`, `docs/goals.md`, the workflows, the acceptance tests, and the data in `docs/preservation.yml`. ## Choose the level by risk and cost - quick (seconds): green items, docs, small fixes. Deterministic checks only. - standard (minutes): code that runs, data files, workflows, anything an executor other than you produced. - deep (longer): releases, security changes, data migrations, red items. Adds live probes and model-based attacks. Run the cheapest level that covers the risk. Raise it when a quick run finds something. ## Run 1. Take a data snapshot before the change when the repository protects data: `python <standard>/scripts/preserve.py snapshot`. 2. Run the verifier: `python <standard>/scripts/verify.py --level <level>`. It runs the standard check, the tests, the data comparison and the scenarios in `scripts/adversarial.py`, and writes a JSON and a Markdown report inside the git folder. 3. Add scenarios for this change. For each category below that applies, write at least one test that tries to make the change fail