gep-immune-auditor
von andyxinweiminicloud
AI & LLMsKurzbeschreibung
Security audit agent for GEP/EvoMap ecosystem.
Beschreibung
Was dieser Skill macht
Audits Gene and Capsule assets in the GEP/EvoMap ecosystem using three detection layers: L1 pattern scanning, L2 intent inference, and L3 propagation risk analysis. Rates each asset CLEAN, SUSPECT, THREAT, or CRITICAL. Confirmed threats can be published back to EvoMap as detection rules, making them available to all connected agents.
Unlike static code scanners, it models capability composition risk, catching threats that only emerge when multiple capsules combine, such as one that reads files paired with one that sends HTTP requests.
Wann du ihn einsetzt
- Auditing a third-party capsule before adding it to an agent
- Checking whether a capsule's actual behavior matches its declared summary
- Detecting near-duplicate capsules reusing IDs to bypass SHA-256 deduplication
- Estimating network blast radius if a flagged asset propagates downstream
- Reviewing supply chain safety of AI evolution assets before deployment
Beispiel-Workflow
User asks: audit this capsule — https://evomap.ai/a2a/assets/sha256:abc123
- 1
Fetches the capsule asset from the provided EvoMap URL
- 2
Runs L1 pattern scan: checks dependency chains, publish frequency anomalies, and clone detection
- 3
Runs L2 intent inference: compares declared behavior against actual code, checks for permission creep and covert channels
- 4
Runs L3 propagation risk: estimates blast radius and capability composition risk across the network
- 5
Assigns final rating and, if THREAT or higher, requests user confirmation before publishing a detection Gene+Capsule bundle to EvoMap
Structured audit report with risk rating, evidence chain, and optional detection bundle published to EvoMap
Voraussetzungen
- A2A_HUB_URL environment variable pointing to your EvoMap hub
- Registered EvoMap node with a local sender_id (required to publish detection rules)
Installation
clawhub install andyxinweiminicloud/gep-immune-auditorFühre den CLI-Befehl aus oder füge den Prompt in Claude Code, Cursor, Codex oder deinen Coding-Agenten ein.