Research

Research seriesL2offensive security

Rethinking Threat Modeling for AI Agents

Classic threat modeling assumes a separable data/control plane and static, checkable trust boundaries. Agentic AI violates both — so a STRIDE-clean threat model can be walked straight through. A red-team analysis of exactly which assumptions break, why, and what replaces per-element enumeration.

Murali Chillakuru·4 articles
  1. 1
    Why Classic Threat Modeling Fails on AI Agents

    A threat model can pass every letter of STRIDE and still be walked straight through. The reason is structural: agents violate the assumptions the method is built on.

  2. 2
    Threat Modeling AI Agents by Capability and Provenance

    If per-element enumeration can't hold the threat, change the unit. Model the agent by the authority it can reach and the trust of the words that steer it.

  3. 3
    Modeling Agent Identity and Delegated Authority

    When an agent acts, three identities are in the room at once. Threat modeling that tracks only one of them misses the delegation chain where agentic authority is actually abused.

  4. 4
    Threat Modeling Multi-Agent and Tool Surfaces

    One agent has a hard-enough trust boundary. Connect several agents and a fleet of tools, and the dangerous surface is no longer any single element — it is the edges between them.