Benchmark Radar Day 73: Auditing Agents Without Reading Their Messages

Day 73 of Benchmark Radar. The scan fetched 1,657 items, published 833, and marked 376 as recommended.

Author: Koutian Wu; GitHub: ktwu01

Benchmark Radar is live. Open the dashboard, star the repository, or read the paper. The Hugging Face dataset provides the data.

ANT tests whether network traffic can reveal what a language-model agent is doing without inspecting private user content. Its 3,114 execution episodes cover 20 tasks, with labels for overall risk, scenario, and basic action. The benchmark gives traffic-based monitors a concrete target, while preserving a boundary around message contents. ANT: A Multi-Granularity Network Traffic Dataset and Benchmark for Agents Behavior Auditing.

FREA evaluates reaction-feasibility verifiers on 751 expert-labeled chemical reactions drawn from several candidate sources. The recorded evaluation found no verifier that led across every source. A filter that works on one generator’s proposals may need a different evaluation when the proposal mix changes. FREA: A Multi-Source Expert Benchmark for Reaction Feasibility Verification.

Why this matters

Choose an evaluation that preserves the information boundary and input mix of the intended workflow. Both can change what a score means.

Scan coverage

This is a keyword-filtered discovery feed, not a representative survey of AI research. The archived briefing used 190 of 1,107 captured evidence records. These findings summarize its cited releases; discovery dates do not necessarily equal original publication dates.

Data and briefing: 2026-10-07 archived snapshot.