Work

FDAA Skill Verification Pipeline v1.0: Defense-in-Depth Security for AI Agent Skills

Year: 2026

Type: article

Abstract: A four-tier verification pipeline for AI agent skills implementing defense-in-depth security. Tier 1 (Fast Pass) provides O(n) pattern matching against known threat signatures. Tier 2 (Guard Model) uses LLM-based semantic analysis to detect Line Jumping, Scope Drift, and Trojan Skills. Tier 3 (Sandbox) monitors runtime behavior in isolated containers. Tier 4 (Registry) provides cryptographic signing with Ed25519 and Merkle tree integrity verification. Validation against the OpenClaw skills ecosystem demonstrates 96.3% detection accuracy for known attack patterns. Reference implementation: fdaa-cli.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.18676239

Authors: Ada, Heydra, Rudi, Substr8 Labs

Language: English

Topic: Adversarial Robustness in Machine Learning

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7130414993

FWCI:

Cites:

Cited by:

Related to: