Efficacité : Autonomous Agent Tool Poisoning & RAG Memory Exfiltration
Sondes et Capteurs Télémétriques Déployés
Section intitulée « Sondes et Capteurs Télémétriques Déployés »| Nom du capteur | Catégorie | Couverture de visibilité |
|---|---|---|
| Hermes Agentic Security Observatory | HOST_AUDIT | Real-time LLM input/output token filtering and tool capability interception. |
| Vector Store Ingestion Telemetry | SIEM | Audit logs for embedding storage in vector database clusters. |
Performances Étape par Étape
Section intitulée « Performances Étape par Étape »| Étape | Nom de l’étape | Détecté | Bloqué | Règle d’alerte déclenchée | Délai |
|---|---|---|---|---|---|
1 | Inert Canary Prompt Injection in MCP Response | ✅ | ⚠️ | LLM-GUARD-T0054: Adversarial Prompt Heuristic Flagged | 8s |
2 | Vector Store Memory Poisoning Injection | ❌ | ⚠️ | NONE | 0s |
3 | Out-of-Band Webhook Exfiltration Probe | ✅ | 🛡️ | EGRESS-FIREWALL-AGENT: Out-of-Band Tool Connection Blocked | 2s |
Zones d’Ombre Défensives Identifiées
Section intitulée « Zones d’Ombre Défensives Identifiées »Directives d’Amélioration Recommandées
Section intitulée « Directives d’Amélioration Recommandées »- Deploy deterministic memory rollback for any agent session flagged by LLM-Guard.
- Enforce strict egress filtering restricting tool network calls to allowlisted API domains.