AASTF

Agentic AI Security Testing Framework — attacks the agent system, not just the model.

Source

Year2026

RoleOpen Source Contributor

LicenseOpen Source

Overview

AASTF tests the full agent execution graph — the LLM plus its tools, memory, and planning loop — against the OWASP Top 10 for Agentic Applications. It intercepts every tool call and planning iteration mid-flight, injects adversarial payloads from 130+ built-in attack scenarios, and produces VULNERABLE / REFUSAL_ECHO / SAFE verdicts with evidence, SARIF output, and CI/CD gating. I contributed new framework adapters and bug fixes to the project.

Highlights

✓130+ OWASP ASI Attack Scenarios

✓Tool-call Interception via LangGraph

✓Three-class Verdicts incl. REFUSAL_ECHO

✓SARIF Reports for GitHub Security Tab

Tech Stack

Python LangGraph OWASP ASI

← All Projects