METR · November 20, 2026

METR: stronger research agents expose the limits of capability benchmarks

Why it matters

METR's latest capability report adds harder tasks to an evaluation suite that frontier models had largely outgrown. Its combined TH1.1 and TH2 results imply a time-horizon doubling period of about 3.6 months, but this is a trend over particular tasks, not a universal measure of autonomous work. Most new task thresholds rely on researcher estimates rather than successful human baselines, and measured performance depends on feedback, scaffolding and scoring choices. Models were strong on coding and optimization with clear success signals, while judgment remained weaker when useful feedback was slow or ambiguous. Model judges also overrated evaluation tasks produced by comparably strong or stronger models. METR's capability index did not show a clear acceleration across the overall public frontier; some developer-specific changes were sensitive to methodology. The operational implication is to evaluate agents on representative, difficult-to-verify work and inspect how they earned their scores. Benchmark saturation and weak human comparisons make confident risk conclusions harder, not unnecessary.

My takeaway: Separate fast-feedback tasks from judgment-heavy work in your evaluation suite. Audit reward hacking and judge calibration, document estimated human baselines, and report task-specific reliability alongside aggregate capability trends.
Keep exploring

More curated notes connected through Model Evaluation and Agent Security.

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.

OWASP GenAI Security Project · guide

OWASP Top 10 for Agentic Applications for 2026

OWASP's community guide organizes agentic-system risk into ten categories, including goal hijacking, tool misuse, identity and privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior. It provides a shared taxonomy and mitigation starting point rather than a certification checklist or evidence that a deployed system is secure.

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.