METR · April 1, 2026

Small fine-tunes change reasoning-trace controllability, not proven monitor evasion

Why it matters

METR fine-tuned four open reasoning models and evaluated the first 240-example checkpoint. Mean instruction compliance on its out-of-distribution CoTControl subset rose from 2.9% to 8.8%. The study did not establish reduced monitorability or generalization to multi-turn agents, and some overlap between training and evaluation instructions remains.

My takeaway: Recheck reasoning-trace controllability after fine-tuning, including held-out instruction types and multiple reasoning budgets. Measure actual monitor detection separately and pair reasoning inspection with action evidence; controllable text alone does not demonstrate successful concealment.
Keep exploring

More curated notes connected through Model Evaluation and AI Red Teaming.

OpenAI News · framework

Pacing model development in an era of cyber-critical capabilities

OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.

OpenAI News · framework

A blueprint for democratic governance of frontier AI

OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.

OpenAI News · framework

OpenAI’s Frontier Governance Framework

OpenAI's 22-page Frontier Governance Framework maps its frontier-model processes to California's Transparency in Frontier AI Act and the EU AI Act's general-purpose AI code. It documents lifecycle risk assessment, cyber-offense and other risk tiers, mitigation and residual-risk decisions, critical-incident handling, security risk management, model reporting, external review, responsibility allocation, and change control.