Research · Virtue AI · 2026
Agentic Red-Teaming on DTAP (DecodingTrust-Agent Platform)
Extending DTAP with new sandboxes and adversarial prompt pipelines that stress-test frontier coding agents.
Working on agentic red-teaming for frontier LLMs on top of DTAP (DecodingTrust-Agent Platform), a controllable simulation platform for AI agent red-teaming across 50+ realistic environments.
Contribution
Generated 500+ adversarial datasets across 10 risk categories, achieving >50% ASR (attack success rate) against Opus 4.7, Llama, and DeepSeek models.
The work focuses on the gap between what red-team benchmarks measure and what actually fails when an agent is dropped into a real environment with tools and state.