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
- When designing or revising halt conditions; use any available CONOPS, threat profile, and data-handling context to identify relevant triggers
- 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:
emergency — Real-incident alert from blue team / SOC
critical — Production data observed in collected evidence
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:
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.
Document schema (derived from the validator's model)
plan/abort.json must validate against AbortPlan. Fields not listed here are not part of the model and are ignored by the validator.
| field |
type |
required |
allowed values / nested keys |
engagement_name |
str |
yes |
|
halt_triggers |
list[AbortTrigger] |
no |
each object: condition, severity, response_action, auto_halt |
abort_signal_channel |
str |
no |
|
recovery_procedure |
str |
no |
|
hallucination_threshold |
int |
no |
>= 1 |
destructive_action_gate |
bool |
no |
|
output_validation |
str |
no |
|
version |
str |
no |
|
last_updated |
str |
no |
|
AbortTrigger object:
| field |
type |
required |
allowed values |
condition |
str |
yes |
|
severity |
TriggerSeverity |
yes |
one of info, warning, critical, emergency |
response_action |
str |
yes |
|
auto_halt |
bool |
no |
|
1---2name: abort-template3description: "Abort / crisis plan generator — halt triggers, response actions, AI-aware safety gates (hallucination threshold, destructive-action gate, output validation)."4allowed-tools: Read Write Edit5metadata:6 source: https://skillmd.com/skills/purpleailab/abort-template7 skillogy:8 version: 19 subdomain: planning10 when_to_use: "create abort plan, emergency halt, crisis procedure, AI safety gate, hallucination guard, destructive action gate"11 tags: abort, halt, crisis, safety, ai-safety, owasp-agentic12 upstream_ref: "Soundwave abort / crisis template — halt triggers, AI-safety gates, hallucination guardrail"13---1415# Abort / Crisis Plan Generator1617The 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.1819AI-specific gates address the failure modes documented in **OWASP Top 10 for Agentic Applications (2025.12)** — hallucination, prompt injection, unverified destructive actions.2021## When to Use2223- When designing or revising halt conditions; use any available CONOPS, threat profile, and data-handling context to identify relevant triggers24- User says "create abort plan", "halt triggers", "emergency stop", "safety gates"2526## Workflow2728### Step 1: Keep the Three Default Triggers2930The `AbortPlan` schema seeds three default halt triggers. Keep all three; do not remove unless the engagement is explicitly read-only:31321. `emergency` — Real-incident alert from blue team / SOC332. `critical` — Production data observed in collected evidence343. `critical` — Scope boundary violation detected3536### Step 2: Add Engagement-Specific Triggers3738Based on the interview answers, add custom `AbortTrigger` entries. Common additions:3940| Engagement signal | Severity | Response |41|---|---|---|42| Tier-3+ adversary emulation against critical infrastructure | `emergency` | Halt + operator-page + 1hr cooldown |43| Compliance-bound data class encountered (GDPR / HIPAA) | `critical` | Halt; redact; await operator |44| Network outage or latency spike on target | `warning` | Pause active scanning; resume after 5min |45| LLM cost-per-objective threshold exceeded | `warning` | Pause; ask operator to approve budget |46| C2 beacon lost for > 30 minutes during active phase | `warning` | Pause; re-establish C2 before continuing |4748### Step 3: Set AI Safety Gates4950- `hallucination_threshold` — default 3 (after 3 unverified-success claims, force evidence-collection mode). Lower (2) for high-stakes engagements; never set 0.51- `destructive_action_gate` — default True. Set False only for read-only / recon-only engagements where no `rm`, `drop`, `format`, `delete-policy` commands will run.52- `output_validation` — default `"verify-evidence-hash"`. Override to `"second-tool-confirm"` for engagements where evidence cannot be hashed (e.g. real-time API responses).5354### Step 4: Set Abort Signal Channel + Recovery Procedure5556- `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.57- `recovery_procedure` — default text covers the standard flow (snapshot → export evidence → run cleanup → operator approval). Override for engagements with custom recovery semantics.5859## Validation Checklist6061Before writing `plan/abort.json`:6263- [ ] At least one `emergency`-severity trigger exists64- [ ] All custom triggers have non-empty `condition` and `response_action`65- [ ] `hallucination_threshold` ≥ 166- [ ] `output_validation` matches one of the documented methods67- [ ] `abort_signal_channel` is set OR explicitly marked "CLI Ctrl+C only"6869## Anti-patterns7071- Removing the EMERGENCY default — guarantees the agent runs through a SOC real-incident alert72- Setting `destructive_action_gate=False` for engagements that include post-exploit or cleanup phases73- Conflating WARNING and CRITICAL — WARNING pauses one step; CRITICAL halts the active objective7475## Output7677Write to `plan/abort.json` validating against `decepticon.core.schemas.AbortPlan`.7879## Document schema (derived from the validator's model)8081`plan/abort.json` must validate against `AbortPlan`. Fields not listed here are not part of the model and are ignored by the validator.8283| field | type | required | allowed values / nested keys |84| --- | --- | --- | --- |85| `engagement_name` | str | yes | |86| `halt_triggers` | list[AbortTrigger] | no | each object: `condition`, `severity`, `response_action`, `auto_halt` |87| `abort_signal_channel` | str | no | |88| `recovery_procedure` | str | no | |89| `hallucination_threshold` | int | no | >= 1 |90| `destructive_action_gate` | bool | no | |91| `output_validation` | str | no | |92| `version` | str | no | |93| `last_updated` | str | no | |9495`AbortTrigger` object:9697| field | type | required | allowed values |98| --- | --- | --- | --- |99| `condition` | str | yes | |100| `severity` | TriggerSeverity | yes | one of `info`, `warning`, `critical`, `emergency` |101| `response_action` | str | yes | |102| `auto_halt` | bool | no | |