Scale AI tests a checker that verifies AI agents' claims of success

In short

Researchers from Scale AI and the University of Arizona report a tool called Sauron, which checks whether AI agents really did what they claim. In tests with deliberately planted failures, the authors say it caught 97% of them and wrongly flagged 2% of clean runs. The results come from benchmark tests, which are standard sets of test tasks, and the paper's summary does not describe any real incident.

What happened

The work was posted on arXiv, a public site for research papers, on 13 October 2026. It looks at how AI agents are graded on long tasks. An agent is an AI system that carries out a series of steps by itself.

Such tasks are often graded by a "judge" using a rubric, which is a checklist of criteria. The authors say the evidence a judge needs can be spread across two places: the state of the environment the agent worked in, and the trace, which is the record of what the agent did. They say a judge may miss this evidence or accept unsupported claims, including an agent's own reports of success.

The authors built Sauron to examine both places. It relies on a claim only after checking it against its source. It also gives each verdict an auditable receipt, which is a record of the supporting evidence.

The authors report three main sets of results:

• Personal-assistant tasks: The tool was tested on 2,586 criteria labelled by people, from 100 task runs. It was compared with four other verifiers, using eight underlying AI models. Sauron had the highest F1 score in all eight. F1 is a measure of accuracy, where 1 is a perfect score.
• Comparison with the closest rival: Using the model GPT-5.6 Sol, Sauron scored 0.96, while the strongest comparison tool, Gandalf, scored 0.92. The authors say Sauron cost 23% less to run in this comparison.
• Planted failures: The authors tested 13 families of faults, alongside clean runs for comparison. Several families simulate reward hacking, which means gaming the scoring system rather than doing the task properly. This is where Sauron caught 97% of failures and flagged 2% of clean runs. The authors say it beat the comparison tools even when those tools were given the full trace.

The authors also used TRAIL, a benchmark for finding and locating real agent errors. They report that Sauron beat both Gandalf and TRAIL's own judge on all three quality measures.

The paper's summary does not say that any real person or organisation was affected. These figures are the authors' own, and the summary does not list open questions or say that anything is still being investigated.

What was done about it

The researchers' response to the problem is the tool itself. The summary does not say whether Sauron is in use outside these tests. Details are in the original report.

Why it matters

The paper links this problem to how AI systems are trained. In reinforcement learning, a system learns from rewards, and a rubric's verdicts can serve as those rewards. The authors say that when this happens, a judge's errors can reinforce undesirable behaviour.