Safety, alignment & governance
IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests
IssueTrojanBench tests whether malicious instructions embedded in software issues can bypass the safeguards of AI coding agents.
Why it matters: For teams using coding agents, the findings caution against treating the agent framework as a strong additional safety net. The abstract points to a need for stronger protections at both the agent and model levels.