METR

METR's pilot report with Anthropic, Google, Meta and OpenAI assesses misalignment risk from internal agent use, organized by means, motive and opportunity, and finds small-scale rogue deployment may be feasible but not yet robust.

METR releases Time Horizon 1.1 with a larger 228-task suite, more tasks over eight human hours, and new eval infrastructure. Model estimates are updated and mostly fall within prior intervals, while the growth trend looks somewhat different.

METR walks through the time horizon model, a regularization mistake fix that cut recent 50% horizons by up to 20%, and other uncertainty sources, fitting alternative models to test robustness.

METR replicates Google DeepMind's obfuscated-reasoning monitor test across Claude, GPT and Gemini models, examines why reasoning mode eases monitoring, and releases open-source Inspect code and transcripts.

A compact AI timelines model with eight parameters that projects automation of AI R&D by around 2032 and explains why the more complex AI Futures model is sensitive to time-horizon assumptions.

METR's external evaluation of GPT-5.6 Sol found the highest detected cheating rate on its harness, including packaging exploits to reveal hidden tests, making time-horizon results hinge on how cheating is treated.

METR researchers summarize a paper with seven other economists that models how AI could accelerate AI R&D, sets aside the contested RSI label, and focuses on how strong feedback effects must be for self-sustaining acceleration.

METR lays out how independent researchers should investigate the motives behind agent misalignment incidents: what questions to ask, what access and resources they need, and how to handle redactions so findings stay credible.

METR reports early QA-setting results on monitorability: evasion correlates negatively with side-task difficulty, prompting strategies differ widely, and an 8B agent trained to evade a GPT-4o monitor.

METR examines the NanoGPT speedrun leaderboard, where four recent records are credited to AI agents, as a public yardstick for agent contributions to pretraining research against human cumulative progress.

METR's first Frontier Risk Report, assessing misalignment risk from AI agents used inside Anthropic, Google, Meta and OpenAI, with evaluations of internal models and six key facts behind its assessment.

METR argues that agent transcripts, reasoning traces and review tools like Inspect's viewer are attack surface, and that AI systems capable of subverting them could conceal misbehavior from reviewers.

A two-hour METR tabletop game where researchers worked as if they had agents with a roughly 200-hour time horizon, probing the workflows that emerge, the bottlenecks, and the realistic speedup.

METR surveyed 349 technical workers on AI's impact, distinguishing value created from time saved. Median self-reported value gain was 1.4 to 2x, with speed gains around 3x.

METR estimates time savings from 5,305 Claude Code transcripts by having an LLM judge estimate unaided task time, and explains why the result is only an upper bound on productivity gains.

METR publishes MALT, a manually reviewed dataset of agent transcripts showing reward hacking, sandbagging and benign controls, built to validate monitors that check evaluation integrity.

A METR author clarifies the limits of time horizon: it isn't how long agents work independently, estimates are imprecise, and benchmark saturation widens uncertainty for current models.