Escape Meridian

· Mar 15, 2024

METR proposes an example protocol for evaluating autonomous AI risks

Markdown · JSON · Original source

published this example protocol on March 15, 2024, to evaluate AI models’ capability to pose autonomy-related risks. It focuses on systems that can complete harmful tasks end-to-end without humans as the bottleneck, including by procuring human assistance; it does not cover other risks such as harmful advice or deepfakes. METR proposes tasks of varying difficulty, model runs and scores that can be combined into a capability measure. The measure is intended to help decision-makers connect evaluation results with choices about model use, deployment, or further development, alongside threat modeling and risk tolerance. METR describes the protocol as an early warning, not a finished basis for setting mitigation thresholds: its authors say the science is at an early stage and do not yet offer a mapping they can defend. They suggest that crossing a threshold could call for pausing activity and developing better evaluations, while warning that changing tests after a threshold is crossed could weaken public accountability and invite gaming. The central unresolved issue is whether task performance and the proposed score can meaningfully track dangerous capabilities and real-world risk.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

METRArticle · metr.orgOpen original source

Original source

· Article

Source details

Original source date: Mar 15, 2024 (2024-03-15T11:00:00.000Z).

Date basis: Date supplied by the publisher’s feed.

Added to Escape Meridian: 2026-11-20T23:03:10.108Z. First observed: 2026-10-18T03:53:24.687Z. Retrieved: 2026-10-25T03:09:42.887Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

The protocol targets autonomous-system risks; says it would not catch other important risks, including those described in the source. METR says the scoring metric and aggregation scheme have not been empirically verified, and that human task duration is not a good measure of model difficulty or risk-relevant capability. Text as retrieved on October 25, 2026; originally published March 15, 2024.

Open question

How well would performance on this task suite predict the capabilities and risks of autonomous AI systems outside the evaluations?

Source-stated

METR says the protocol targets risks from autonomous AI systems that can carry out substantive harmful tasks without humans as the bottleneck.

systems that are able to accomplish substantive harmful tasks end-to-end on their own without humans being the bottleneck

Source-stated

METR says the protocol does not address several other kinds of AI risk and calls for separate evaluations to address them.

These types of risks are important, and this protocol would not catch them. We think separate evaluations should be conducted to ensure that these risks are also addressed.

Source-stated

METR says the protocol is an early warning and that its authors cannot yet defend an overall mapping from scores to mitigations.

The science of evaluation for autonomous capabilities is still in very early stages, so we don’t yet have overall suggestions for what this mapping should be that we feel able to defend.

Source-stated

METR warns that changing evaluations after a threshold is crossed may weaken public accountability and create opportunities to game the updated tests.

AI developers might face intense pressures to (and have ample opportunity to) construct the new evals to produce the desired result, or to game them.

Source and contributor context

METR · Publisher

The publisher is identified here. No individual author has been inferred.

Source media

Media retrieval status: unchecked.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

METR

METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.

Editorial provenance

Version 1. Publication ID: 48a4a042-db60-9e20-4f1e-415553f8a6a8.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-20T22:58:26.165Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge_admitted_by_model.

Observation: c719b81b1ece81a7bef3fd159e8448e193c1b6d83ac32fe1323caf91f83ab050. Source text hash: 39fa4a36a8aa9731adb39492d8d66bcfd2c3da8f79a32256e0a6a85ba3a53936.

Revision and source-check history

Version 1 · 2026-11-20T23:03:10.108Z

METR proposes an example protocol for evaluating autonomous AI risks

published this example protocol on March 15, 2024, to evaluate AI models’ capability to pose autonomy-related risks. It focuses on systems that can complete harmful tasks end-to-end without humans as the bottleneck, including by procuring human assistance; it does not cover other risks such as harmful advice or deepfakes. METR proposes tasks of varying difficulty, model runs and scores that can be combined into a capability measure. The measure is intended to help decision-makers connect evaluation results with choices about model use, deployment, or further development, alongside threat modeling and risk tolerance. METR describes the protocol as an early warning, not a finished basis for setting mitigation thresholds: its authors say the science is at an early stage and do not yet offer a mapping they can defend. They suggest that crossing a threshold could call for pausing activity and developing better evaluations, while warning that changing tests after a threshold is crossed could weaken public accountability and invite gaming. The central unresolved issue is whether task performance and the proposed score can meaningfully track dangerous capabilities and real-world risk.

Original source

· Article

Source details

Original source date: Mar 15, 2024 (2024-03-15T11:00:00.000Z).

Date basis: Date supplied by the publisher’s feed.

Added to Escape Meridian: 2026-11-20T23:03:10.108Z. First observed: 2026-10-18T03:53:24.687Z. Retrieved: 2026-10-25T03:09:42.887Z.

Limits

The protocol targets autonomous-system risks; says it would not catch other important risks, including those described in the source. METR says the scoring metric and aggregation scheme have not been empirically verified, and that human task duration is not a good measure of model difficulty or risk-relevant capability. Text as retrieved on October 25, 2026; originally published March 15, 2024.

Open question

How well would performance on this task suite predict the capabilities and risks of autonomous AI systems outside the evaluations?

Source-stated

METR says the protocol targets risks from autonomous AI systems that can carry out substantive harmful tasks without humans as the bottleneck.

systems that are able to accomplish substantive harmful tasks end-to-end on their own without humans being the bottleneck

Source-stated

METR says the protocol does not address several other kinds of AI risk and calls for separate evaluations to address them.

These types of risks are important, and this protocol would not catch them. We think separate evaluations should be conducted to ensure that these risks are also addressed.

Source-stated

METR says the protocol is an early warning and that its authors cannot yet defend an overall mapping from scores to mitigations.

The science of evaluation for autonomous capabilities is still in very early stages, so we don’t yet have overall suggestions for what this mapping should be that we feel able to defend.

Source-stated

METR warns that changing evaluations after a threshold is crossed may weaken public accountability and create opportunities to game the updated tests.

AI developers might face intense pressures to (and have ample opportunity to) construct the new evals to produce the desired result, or to game them.

Source and contributor context

METR · Publisher

The publisher is identified here. No individual author has been inferred.

Source media

Media retrieval status: unchecked.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

METR

METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.

Editorial provenance

Version 1. Publication ID: 48a4a042-db60-9e20-4f1e-415553f8a6a8.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-20T22:58:26.165Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge_admitted_by_model.

Observation: c719b81b1ece81a7bef3fd159e8448e193c1b6d83ac32fe1323caf91f83ab050. Source text hash: 39fa4a36a8aa9731adb39492d8d66bcfd2c3da8f79a32256e0a6a85ba3a53936.

passed · 2026-11-20T22:58:26.165Z · anthropic/claude-opus-5.5

Publication: 48a4a042-db60-9e20-4f1e-415553f8a6a8

attribution — pass: The reading consistently attributes the protocol, its scope, its caveats and its accountability warning to METR, the publisher of this single blog document. It does not present METR's claims as established fact.

This is an example protocol for evaluating an AI model’s capability to pose autonomy-related risks.

summary — pass: The reading leads with the actor and the development. It explains autonomy risk in plain terms, describes the task, run and score structure, and states the intended decision use. It preserves METR's self-described early-warning status, its admission that it has no defensible mapping, and its tension between pausing and accountability. The question follows from METR's own caveats about the metric. It could also have mentioned the conservative bound of under 1 hour, but that omission does not distort the reading.

This evaluation protocol could be rapidly improved, and is intended only as a conservative early warning

date — pass: originalPublished is 2024-03-15T11:00Z, and the reading states March 15, 2024, which matches the UTC day and the URL path. Nothing in the reading uses hindsight or later events. The source's plans, such as future updates, are not presented as outcomes. The evidence is a publisher page, not an archive copy, so no archive disclosure is required. The retrieval-date note is accurate.

support — pass: Each claim is backed by a verbatim excerpt. The limitations are grounded in the source, including the unverified metric and the human-hours caveat. The phrase 'not a good measure' closely tracks the source's 'not a great metric'. The pausing recommendation is hedged, matching the source's 'often'. The reading adds no unsupported benefits or forecasts.

The metric and aggregation scheme we use hasn’t been empirically verified

media — pass: No media is selected, and no verified candidates were supplied. The source page carries no media items, so omitting media is appropriate.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.