Megan Kinniment

Megan Kinniment is a researcher on METR's technical staff who builds benchmarks for measuring what AI agents can do on their own. Trained in physics, with a Master's degree from the University of Oxford, she worked with the Center on Long-Term Risk, the Future of Humanity Institute and 1DaySooner, and by 2023 she was on the Alignment Research Center's evaluations project. In December 2023, she was first author of a report testing language-model agents on tasks tied to autonomous replication and adaptation; she went on to coauthor RE-Bench, HCAST and a 2025 paper finding that AI time horizons were doubling about every seven months. Her 2026 work with colleagues introduced SHUSHCAST, a prototype test of whether monitor models catch agents doing hidden side tasks. She leads METR's benchmark creation effort and mentors MATS fellows on monitorability and alignment evaluations.

Now at: METR
Works: 7
Citations: 65
h-index: 4

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

By 2022

A physics Master's at Oxford

She studies physics, doing malaria diagnostics research at university, and holds a Master's degree from the University of Oxford. She works with the Center on Long-Term Risk, the Future of Humanity Institute and 1DaySooner, and a CLR grant funds her independent projects, including ML interpretability. In June 2022, she posts "Exploring Mild Behaviour in Embedded Agents", which proposes inducing mild optimisation in embedded agents through their utility functions.
2022

MATS and GPT-2 interpretability

In summer 2022, she is a model evaluations fellow in MATS's first full-scale cohort of 31 fellows. Her October 2022 post "Recall and Regurgitation in GPT2" uses causal tracing to compare how GPT2-XL completes cached phrases and factual statements. In December 2022, she and Evan R. Murphy publish tests for non-myopic steering behaviour in InstructGPT models, with Kinniment doing most of the experiment design.
December 2023

Autonomous replication evaluations

Working on the Alignment Research Center's evaluations project, she is first author of "Evaluating Language-Model Agents on Realistic Autonomous Tasks" in December 2023. The report tests four example agents on 12 tasks tied to autonomous replication and adaptation, and the agents complete only the easiest ones. With Beth Barnes, she posts a December 2023 bounty for hard tasks for LLM agents at METR, formerly ARC Evals.
November 2024

The RE-Bench benchmark

She coauthors RE-Bench, released in November 2024, which compares AI agents with 61 human experts on 7 open-ended ML research engineering environments. Given 2 hours per environment, the best AI agents score 4 times higher than the experts.
March 2025

HCAST and AI time horizons

In March 2025, she coauthors HCAST, a benchmark of 189 software, cybersecurity, ML engineering and reasoning tasks calibrated with 563 human baselines. The paper "Measuring AI Ability to Complete Long Software Tasks", which she coauthors in March 2025, introduces the 50%-task-completion time horizon and finds it doubling about every seven months since 2019. Her METR essay of March 2025 argues that legible and faithful AI reasoning makes safety problems easier to catch.
July 2025

An IEEE Spectrum interview

In a July 2025 IEEE Spectrum interview about the time-horizon paper, she says she personally did not expect so clear an exponential trend. If models reach month-long tasks, she says, the big capability to watch is accelerating AI research and development itself.
January 2026

SHUSHCAST monitorability evaluations

In January 2026, she and four colleagues publish METR's early work on monitorability evaluations, presenting SHUSHCAST, a prototype that tests whether monitors catch AI agents secretly doing side tasks. With OpenAI's reasoning traces, monitor catch rates against GPT-5 rise by over 50 percentage points.
As of October 2026

Technical staff at METR

She is a member of technical staff at METR, where she leads the benchmark creation effort and has also led task development and the monitoring stream. She mentors a MATS Summer 2026 stream on monitorability and alignment evaluations. For METR's Frontier Risk Report of May 2026, a pilot assessment of misalignment risks from AI agents inside Anthropic, Google, Meta and OpenAI, she works on evaluations and infrastructure.