Rohan Paul (@rohanpaul_ai)T2·78 pts
- BadEvolve needs only 20 LLM-optimized poisoned tasks to bake backdoor rules into reusable agent skills
- 79.6% payload injection rate on Kimi-K2.5 across 4 frameworks and 3 benchmarks, versus 37.6% for the best baseline
- Clean-task accuracy fell just 1.1 points, making the backdoor nearly invisible to standard evaluation
- Defense: treat external or user-submitted learning tasks as untrusted and review learned skills before reuse