AHA is an automated red-teaming framework that uses an agentic research environment to discover reusable vulnerability knowledge about another production-grade LLM agent. It proposes vulnerability hypotheses, builds validators, instantiates attacks, executes in a sandbox, reflects on trajectories, and promotes confirmed findings into a Vulnerability Concept Graph (VCG). In three scenarios (including direct and indirect attacks) on Claude Code and Codex, discovered concepts reveal reusable vulnerability cores.
AHA is an automated red-teaming framework that discovers reusable vulnerability knowledge about another production-grade LLM agent through an agentic research environment. It employs a falsifiable discovery loop, promoting confirmed vulnerabilities into a Vulnerability Concept Graph (VCG). In three scenarios on Claude Code and Codex, discovered concepts reveal reusable vulnerability cores.
AHA's Vulnerability Concept Graph (VCG) links attack surfaces to unsafe trajectories through claims, enabling conditions, validators, migration predictions, and evidence, potentially becoming a standard method for automated red-teaming knowledge reuse.
This work indicates that security testing of production-grade agents is shifting from one-off attacks to systematic discovery and reuse of vulnerability knowledge, potentially driving standardization of agent security evaluation.
For companies deploying production-grade agents, AHA provides an automated red-teaming method that may reduce security testing costs and improve vulnerability discovery efficiency.
Verifiable next signal: AHA framework is applied to more production-grade agents (e.g., GPT-4 agents or open-source agents), or VCG is integrated into CI/CD security testing pipelines.