abort-template

# Abort / Crisis Plan Generator The abort plan defines **when the agent stops**. Without it, the autonomous loop runs through real-incident alerts, scope violations, and production-data exposures as if they were normal operations. AI-specific gates address the failure modes documented in **OWASP Top 10 for Agentic Applications (2025.12)** — hallucination, prompt injection, unverified destructive actions. ## When to Use - After CONOPS + Threat Profile + Data Handling are written (the kill chain shapes which triggers are relevant) - User says "create abort plan", "halt triggers", "emergency stop", "safety gates" ## Workflow ### Step 1: Keep the Three Default Triggers The `AbortPlan` schema seeds three default halt triggers. Keep all three; do not remove unless the engagement is explicitly read-only: 1. `EMERGENCY` — Real-incident alert from blue team / SOC 2. `CRITICAL` — Production data observed in collected evidence 3. `CRITICAL` — Scope boundary violation detected ### Step 2: Add Engagement-Specific Triggers Based on the interview answers, add custom `AbortTrigger` entries. Common additions: | Engagement signal | Severity | Response | |---|---|---| | Tier-3+ adversary emulation against critical infrastructure | `EMERGENCY` | Halt + operator-page + 1hr cooldown | | Compliance-bound data class encountered (GDPR / HIPAA) | `CRITICAL` | Halt; redact; await operator | | Network outage or latency spike on target | `WARNING` | Pause active scanning; resume after 5min | | LLM cost-per-objective threshold exceeded | `WARNING` | Pause; ask operator to approve budget | | C2 beacon lost for > 30 minutes during active phase | `WARNING` | Pause; re-establish C2 before continuing | ### Step 3: Set AI Safety Gates - `hallucination_threshold` — default 3 (after 3 unverified-success claims, force evidence-collection mode). Lower (2) for high-stakes engagements; never set 0. - `destructive_action_gate` — default True. Set False only for read-only / recon-only engagements where no `rm`, `drop`, `format`, `delete-policy` commands will run. - `output_validation` — default `"verify-evidence-hash"`. Override to `"second-tool-confirm"` for engagements where evidence cannot be hashed (e.g. real-time API responses). ### Step 4: Set Abort Signal Channel + Recovery Procedure - `abort_signal_channel` — operator's out-of-band channel for forcing halt (HTTP endpoint, file marker, ask_user_question). Leave empty only if Ctrl+C-via-CLI is the only abort path. - `recovery_procedure` — default text covers the standard flow (snapshot → export evidence → run cleanup → operator approval). Override for engagements with custom recovery semantics. ## Validation Checklist Before writing `plan/abort.json`: - [ ] At least one `EMERGENCY`-severity trigger exists - [ ] All custom triggers have non-empty `condition` and `response_action` - [ ] `hallucination_threshold` ≥ 1 - [ ] `output_validation` matches one of the documented methods - [ ] `abort_signal_channel` is set OR explicitly marked "CLI Ctrl+C only" ## Anti-patterns - Removing the EMERGENCY default — guarantees the agent runs through a SOC real-incident alert - Setting `destructive_action_gate=False` for engagements that include post-exploit or cleanup phases - Conflating WARNING and CRITICAL — WARNING pauses one step; CRITICAL halts the active objective ## Output Write to `plan/abort.json` validating against `decepticon.core.schemas.AbortPlan`.