IAPS report recommends frontier labs disclose chain-of-thought monitorability indicators

1. 20:31
   IAPSSelected

   IAPS report: transparency disclosures for chain-of-thought monitorability

   IAPS released a report arguing that frontier AI developers should regularly disclose a set of chain-of-thought (CoT) monitorability metrics so that independent evaluators can track their trajectories of change. The report selected five metrics from 21 candidates: third-party chain-of-thought access and infrastructure specifications, naturalness, faithfulness, detection (the proportion of harmful behavior that escapes monitoring systems), and Uplift (the performance gain that chain-of-thought access provides to monitoring). An audit of OpenAI, Anthropic, and Google DeepMind showed that none reported all the metrics, but each metric had been reported by at least one, so complete cross-institution reporting is feasible. The report cites recent disclosures from the labs saying that monitorability is declining, for example the GPT-6 Astra system card released in September 2026 shows that its chain-of-thought monitorability dropped substantially compared with the previous generation. The report recommends incorporating these metrics into existing System Card and safety evaluation practices, and conducting deeper audits by embedded evaluators or independent verification organizations; the first version of the report was a preprint from July 2026 and has been accepted by the ICML 2026 Technical AI Governance Workshop.

2. 18:33
   arXiv: Alignment, Deception & Oversight

   Transparency Disclosures for Chain-of-Thought Monitorability

   The paper argues that frontier AI developers should disclose specific metrics so that threats to chain-of-thought (CoT) monitoring are trackable and so that parties can coordinate to protect CoT monitorability. The authors identify five metrics corresponding to the five-step chain on which CoT monitoring depends, covering third-party CoT access and infrastructure specifications, CoT quality (naturalness and faithfulness), and monitoring effectiveness (detection and capability uplift), with the first also supporting external validation of the remaining metrics. An audit of current disclosures by Anthropic, OpenAI and Google DeepMind shows that no developer covers all five, but each metric has been reported by at least one frontier developer. Given that existing disclosures rely mainly on developer self-assessment, the paper recommends that resident evaluators or independent verification organizations repeat cross-developer audits with deeper access.

Current attention 39·Peak in the comparable period 43 (Nov 11 14:00)·Change over the comparable past 24 hours -3%