Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
To evaluate the expanding attack surface of autonomous tool-using agents, the authors built SAGE-RT, a black-box red-teaming framework guided by a seven-domain risk taxonomy that operates purely on high-level system descriptions. The pipeline automatically synthesizes 120 multi-turn adversarial scenarios per domain to systematically probe multi-step execution and inter-agent coordination without privileged access. Evaluating CrewAI and AutoGen architectures across four base models uncovered critical vulnerabilities, including an 85% behavioral failure rate, 56.25% governance risk, and 65% privacy leakage in multi-agent environments.
Autonomous multi-agent systems leak private data in 65% of runs and hit behavioral failure rates up to 85% under fully automated, black-box adversarial probing.
Agentic systems are rapidly moving to production, where they read untrusted inputs, call tools with real permissions, and act autonomously, expanding the security surface beyond chat-only models. Yet standard evaluations remain single-turn and fail to capture multi-step agent vulnerabilities. We present a systematic black-box framework for risk-aware agent evaluation requiring only basic system descriptions. Our approach introduces: (1) a seven-domain taxonomy mapping observable behaviors to risk categories, (2) fully automated SAGE-RT red teaming producing 120 adversarial scenarios per domain, and (3) human-validated evaluation using LLM judges. Empirical validation across two agent architectures (CrewAI and AutoGen) with four base models reveals alarming patterns: 56.25\% average governance risk, 65\% privacy risk in multi-agent configurations, and agent behavior vulnerabilities reaching 85\%. Our black-box approach effectively identifies critical architectural vulnerabilities without privileged access, providing a scalable path toward safer agent deployments.