Reproduce “ASTRA: Autonomous Spatial-Temporal Red-teaming for AI Software Assistants” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiangzhe Xu, Guangyu Shen, Zian Su, Siyuan Cheng, Hanxi Guo, Lu Yan, Xuan Chen, Jiasheng Jiang and 4 more

Researchers at Purdue University developed ASTRA, an automated multi-agent system for red-teaming AI coding assistants by adaptively exploring both input space and reasoning paths to find safety vulnerabilities in realistic scenarios. This system uncovered 11-66% more issues and generated 17% more effective alignment training cases compared to previous methods.

Key insights

  • Realistic inputs and systematic exploration of both input space (spatial) and model reasoning processes (temporal) are crucial for uncovering meaningful and actionable AI safety vulnerabilities in coding assistants.
  • Existing blue-teaming defenses, while achieving high Defense Success Rates, often compromise model utility (e.g., Circuit Breaker's over-generalization) or exhibit brittle reasoning paths that can be exploited (e.g., Deliberative Alignment).
  • Identifying and exploiting discrepancies in a model's Chain-of-Thought explanations can effectively reveal underlying reasoning flaws, even when initial outputs appear safe.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.