harden-and-remediate-ai-systemlisted

# Skill: harden-and-remediate-ai-system > **Invoked by:** `adversarial-testing-engineer` (primary). Also consulted by `ai-redteam-lead` to confirm each in-scope risk has a plausible remediation before the plan tests it. > > **When to invoke:** "What do we fix and how?"; "harden this against prompt injection"; "stop the agent from unauthorized tool calls"; "add guardrails"; any move from findings to a hardened, retested system. > > **Output:** triaged findings (severity, dedup, reproducibility) + a defense-in-depth remediation plan (layered controls) + a retest of each fix with the exact attack + the attack baked into the regression harness. ## Procedure 1. **Triage the findings by likelihood × impact.** Dedup, confirm each is reproducible, assign severity on the plan's scale, and rank P0→P2. A ranked, reproducible work queue — not a wall of transcripts. Impact rises with what the system can *do* and how many it affects; likelihood with reachability (public/unauthenticated/no-rate-limit). 2. **Remediate defense-in-depth — never a single filter.** For each finding, layer controls from [`../../knowledge/ai-red-teaming-patterns-2026.md`](../../knowledge/ai-red-teaming-patterns-2026.md): - **Input guardrails** — classify/sanitize untrusted input, detect known injection patterns (a *layer*, bypassable alone). - **Output guardrails** — scan responses for leaked secrets/PII/policy violations before they reach the user or a downstream sink. - **Injection-resistant prompt s