Defining the Agentic Attack Surface
Published: Draft
Traditional security tooling is optimized for static infrastructure and human operator workflows. Autonomous AI agents introduce a dynamic, adaptive operational layer that mutates faster than signature-based or rule-centric detection can follow.
The Agentic Attack Surface encompasses vulnerabilities emerging from multi-agent orchestration, delegated tool execution, memory/state manipulation, and policy boundary erosion over iterative task cycles.
The Four Core Domains
- Multi-Agent Collusion – Coordinated behavior that bypasses guardrails through distributed intent.
- Malicious Tool Use – Legitimate capabilities repurposed for exfiltration or privilege escalation.
- AI Memory Poisoning – Persistent manipulation of long-term state to shift agent objectives.
- Insider Agent-as-Proxy – Leveraging trusted internal agents to execute unapproved actions.
Securing this layer requires continuous semantic observation, autonomous adversarial simulation, and policy-aware execution governance. This is why we built the ASFM.