Attack Surface Enumeration
Security & IdentityDescription
Adversary systematically maps agent's tool definitions, capability boundaries, activation triggers, and embedded knowledge to enable targeted attacks. Reconnaissance phase for more sophisticated exploits.
Attacker systematically queries agent to enumerate its available tools, system prompt structure, and knowledge base contents before launching a targeted attack.
Membership inference and model inversion have been demonstrated in peer-reviewed research against real ML services, and verbatim memorisation of training data has been shown in large language models including GPT-2. Confirmed exploitation of these techniques to extract personal data from production systems at scale has not been publicly documented.
Primary mitigations
- Minimise configuration exposure
- capability disclosure controls
- detect enumeration patterns in agent interactions.
Detection signals
Configuration enumeration attempt detection; tool definition disclosure events.