Benchmark
AgentDyn
A benchmark framework designed to evaluate the dynamic task execution capabilities of LLM agents. It focuses on task success rates in multi-step, tool-calling-chain scenarios and supports the injection of adversarial attacks such as prompt injection to assess the effectiveness of defensive measures.
Timeline (last 90 days)
Sep 30
In the AgentDyn benchmark, Environment Steering's task success rate exceeds that of the no-defense baseline.
Unverified50%
Sep 30
In the AgentDyn benchmark, environmental guidance reduced the attack success rate to 0%.
Unverified50%