Agent research is treating scores and tool actions as audit targets
The day’s strongest signal is auditability for agentic systems. BenchJack attacks benchmark harnesses before agents run. Rollout Cards asks evaluations to publish rollout evidence.
Topic summary
Benchmark Security
Inspectable Agent Operations
Agent teams now have concrete audit gates to copy: pre-release reward-hacking runs for benchmarks, centrally governed MCP servers with trace requirements, and path-level review for code translation.