A new evaluation framework called AgentAudit scores AI agents on trustworthiness, not just whether they finish a job.
Researchers built AgentAudit to audit the full execution trace of an AI agent, covering planning, tool selection, tool execution, memory and reasoning, rather than checking a single stage like most existing tools. It attaches to an agent without altering how it runs, then scores the trace across ten dimensions including instruction integrity, tool correctness and security. The team ran five models, OpenAI GPT-5, Claude Sonnet 5, Sarvam 105B, Llama 3.3 70B and Gemini 2.5 Flash, through nine capability and adversarial tasks. Claude Sonnet 5 topped the resulting Composite Trust Score at 95.1 out of 100, GPT-5 followed at 80.6, and Sarvam 105B, Llama 3.3 70B and Gemini 2.5 Flash trailed at 57.6, 45.7 and 22.6.
The gap that matters isn't just the score spread, it's what's behind it: several non-frontier models were repeatedly flagged as Unsafe_Compliance on adversarial tasks, meaning they went along with harmful instructions rather than simply failing to complete them. A pass or fail benchmark would never catch that distinction, which is the whole pitch behind tools like AgentBench, AgentDojo and ASB that AgentAudit is trying to supersede by grading the entire pipeline instead of one slice of it.
One catch worth flagging: the traces were graded by a single judge model that was itself one of the five being tested, the kind of self-grading setup that tends to earn an asterisk in any trust evaluation.