AFB Taxonomy Specification

The Agent Failure Boundary (AFB) taxonomy is an open security specification for agentic AI systems.

It defines four invariant failure boundaries in the agentic execution loop:

• AFB01 - Context Poisoning: the model ingests corrupted, forged, or manipulated context.
• AFB02 - Model Boundary Compromise: integrity/confidentiality failures at the model input/output boundary.
• AFB03 - Instruction Hijack: model output becomes unsafe instructions for the agent layer.
• AFB04 - Unauthorized Action: the agent attempts or performs an action outside authorized policy.

Who this is for

• Security engineers designing controls for AI agents.
• Agent builders implementing policy and enforcement boundaries.
• Researchers analyzing structural failure modes in autonomous systems.

How to use this spec

1. Start with spec/afb-v1.md and spec/afb-v2.md for the full taxonomy texts derived from the source papers.
2. Use boundary definitions to map architecture risks to loop transitions (Context -> Model -> Agent -> Act).
3. Use owasp-mapping.md to align AFB categories with OWASP LLM and OWASP Agentic categories (interpretive overlap only).
4. Use the examples in examples/ to see concrete AFB01 and AFB04 exposure patterns.

Project links

• https://plarix.dev