METR · November 16, 2026

METR: check whether your alignment evaluations reveal what they test

Why it matters

METR places investigator agents inside 19 alignment and 36 capability-evaluation environments and asks what each task is testing. It compares direct probability estimates with open-ended explanations graded by an LLM, averaging three runs per environment. On this dataset, Opus 5.5 ranked every alignment environment above every capability environment with both methods. Explanations pointed to task-level clues such as implausible stakes, exploitable graders and conflicting objectives, alongside environment-level clues such as obviously simulated infrastructure. This is evidence that the selected evaluations are distinguishable under interrogation, not proof that models spontaneously recognize every test or change their behavior because of that recognition. The questions themselves can prime evaluation awareness, and METR explicitly does not measure the causal effect on task behavior. The note also received less review than METR's formal research posts. Its proposed use is iterative evaluation design: identify revealing clues, revise the scenario and measure again.

My takeaway: Run both open-ended and direct investigator probes against safety scenarios and ordinary capability tasks. Track discrimination and calibration, remove unintended simulation clues, and separately test whether any revision changes actual behavior before interpreting a lower recognition score as stronger safety evidence.
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.

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.