Escape Meridian

Ajeya Cotra · Nov 21, 2024

Luca Righetti argues in Planned Obsolescence that OpenAI’s CBRN tests leave a key question open

Markdown · JSON · Original source

Luca Righetti’s article was published on November 21, 2024. He argues that ’s tests do not clearly establish whether o1-preview can meaningfully help a novice create a CBRN threat. CBRN means chemical, biological, radiological, and nuclear. The distinction matters because OpenAI’s voluntary safety commitments tied a “high” risk finding to safeguards before release, while its “medium” level did not carry the same commitment. Righetti says OpenAI was developing a hands-on laboratory test but released the model before it concluded, without sharing preliminary results; instead, it relied on three multiple-choice tests as proxies. He examines their limits: one result was comparable to experts using Google, another lacked a reported human comparison, and the test showing a gap from experts did not give o1-preview the specialized tools experts used. He also notes that novices might work with a model interactively, unlike a model taking a test alone. Righetti credits OpenAI’s Preparedness Team for publishing enough detail to scrutinize its claims, and welcomes its ongoing wet lab study. His conclusion is that the reported proxy tests leave him unable to tell whether o1-preview meets OpenAI’s definition of “high risk”—and unsure whether OpenAI can tell either. The evidence he discusses does not include results from the unfinished hands-on laboratory test.

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.

Ajeya CotraArticle · planned-obsolescence.orgOpen original source

Original source

Ajeya Cotra · Article

Source details

Original source date: Nov 21, 2024 (2024-11-21T15:56:48+00:00).

Date basis: Date stated on the publisher’s page.

Added to Escape Meridian: 2026-11-16T12:18:24.763Z. First observed: 2026-10-30T14:00:39.681Z. Retrieved: 2026-10-30T16:35:45.162Z.

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

Text as retrieved on October 30, 2026; originally published November 21, 2024. The hands-on laboratory test was still in development in the article; Righetti reports no results from it. Righetti discloses that Open Philanthropy, his employer, funded the ProtocolQA and Cloning Scenarios .

Open question

What would results from ’s hands-on laboratory test show about whether o1-preview can meaningfully help novices with laboratory work?

Source-stated

Righetti argues that OpenAI’s proxy test results do not clearly establish whether o1-preview can meaningfully help a novice create a CBRN threat.

OpenAI says o1-preview can't meaningfully help novices make chemical and biological weapons. Their test results don’t clearly establish this.

Source-stated

Righetti says OpenAI was developing a hands-on laboratory test, but released o1-preview before it concluded and did not share preliminary results.

they released o1-preview before that test concluded and didn’t share any preliminary results.

Source-stated

Righetti says OpenAI used three multiple-choice tests as proxies for laboratory help and questions whether their results support the claim about novices.

Instead, they cite three multiple-choice tests as proxies for laboratory help.

Source-stated

Righetti says the risk distinction mattered because OpenAI’s voluntary commitments required extra safeguards before release at “high” risk, but not at “medium” risk.

At the medium risk level, OpenAI didn't commit to doing anything special to make o1-preview safe.

Source-stated

Righetti says OpenAI’s Preparedness Team published enough detail to let him scrutinize its safety claims, while his conclusion about the high-risk threshold remains unresolved.

I can't tell if o1-preview meets OpenAI's definition of "high risk." And I don't know if OpenAI can tell either.

Source and contributor context

Ajeya Cotra · 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.

Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

OpenAI

OpenAI develops AI models and products and publishes research, release announcements and evaluations.

Editorial provenance

Version 1. Publication ID: 31213c90-84ed-f21b-7ef4-709d58c83669.

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

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-16T12:13:40.531Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.

Observation: 950c024d85ef593caa8185e4ae66c8bdf5ce6bcb2b18f21099126dac7cdc9439. Source text hash: 178b7696949c44a0dc8dbc564bf14ee2ccd689178110e71e3cae43dd64c6d0e3.

Revision and source-check history

Version 1 · 2026-11-16T12:18:24.763Z

Luca Righetti argues in Planned Obsolescence that OpenAI’s CBRN tests leave a key question open

Luca Righetti’s article was published on November 21, 2024. He argues that ’s tests do not clearly establish whether o1-preview can meaningfully help a novice create a CBRN threat. CBRN means chemical, biological, radiological, and nuclear. The distinction matters because OpenAI’s voluntary safety commitments tied a “high” risk finding to safeguards before release, while its “medium” level did not carry the same commitment. Righetti says OpenAI was developing a hands-on laboratory test but released the model before it concluded, without sharing preliminary results; instead, it relied on three multiple-choice tests as proxies. He examines their limits: one result was comparable to experts using Google, another lacked a reported human comparison, and the test showing a gap from experts did not give o1-preview the specialized tools experts used. He also notes that novices might work with a model interactively, unlike a model taking a test alone. Righetti credits OpenAI’s Preparedness Team for publishing enough detail to scrutinize its claims, and welcomes its ongoing wet lab study. His conclusion is that the reported proxy tests leave him unable to tell whether o1-preview meets OpenAI’s definition of “high risk”—and unsure whether OpenAI can tell either. The evidence he discusses does not include results from the unfinished hands-on laboratory test.

Original source

Ajeya Cotra · Article

Source details

Original source date: Nov 21, 2024 (2024-11-21T15:56:48+00:00).

Date basis: Date stated on the publisher’s page.

Added to Escape Meridian: 2026-11-16T12:18:24.763Z. First observed: 2026-10-30T14:00:39.681Z. Retrieved: 2026-10-30T16:35:45.162Z.

Limits

Text as retrieved on October 30, 2026; originally published November 21, 2024. The hands-on laboratory test was still in development in the article; Righetti reports no results from it. Righetti discloses that Open Philanthropy, his employer, funded the ProtocolQA and Cloning Scenarios .

Open question

What would results from ’s hands-on laboratory test show about whether o1-preview can meaningfully help novices with laboratory work?

Source-stated

Righetti argues that OpenAI’s proxy test results do not clearly establish whether o1-preview can meaningfully help a novice create a CBRN threat.

OpenAI says o1-preview can't meaningfully help novices make chemical and biological weapons. Their test results don’t clearly establish this.

Source-stated

Righetti says OpenAI was developing a hands-on laboratory test, but released o1-preview before it concluded and did not share preliminary results.

they released o1-preview before that test concluded and didn’t share any preliminary results.

Source-stated

Righetti says OpenAI used three multiple-choice tests as proxies for laboratory help and questions whether their results support the claim about novices.

Instead, they cite three multiple-choice tests as proxies for laboratory help.

Source-stated

Righetti says the risk distinction mattered because OpenAI’s voluntary commitments required extra safeguards before release at “high” risk, but not at “medium” risk.

At the medium risk level, OpenAI didn't commit to doing anything special to make o1-preview safe.

Source-stated

Righetti says OpenAI’s Preparedness Team published enough detail to let him scrutinize its safety claims, while his conclusion about the high-risk threshold remains unresolved.

I can't tell if o1-preview meets OpenAI's definition of "high risk." And I don't know if OpenAI can tell either.

Source and contributor context

Ajeya Cotra · 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.

Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

OpenAI

OpenAI develops AI models and products and publishes research, release announcements and evaluations.

Editorial provenance

Version 1. Publication ID: 31213c90-84ed-f21b-7ef4-709d58c83669.

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

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-16T12:13:40.531Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.

Observation: 950c024d85ef593caa8185e4ae66c8bdf5ce6bcb2b18f21099126dac7cdc9439. Source text hash: 178b7696949c44a0dc8dbc564bf14ee2ccd689178110e71e3cae43dd64c6d0e3.

passed · 2026-11-16T12:13:40.531Z · anthropic/claude-opus-5.5

Publication: 31213c90-84ed-f21b-7ef4-709d58c83669

attribution — pass: The source byline names Luca Righetti. The reading attributes the argument to him and places it on the Planned Obsolescence site, which matches the URL. OpenAI's claims stay attributed to OpenAI, and Righetti's judgments stay attributed to him.

Luca Righetti Nov 21, 2024

summary — pass: The reading leads with Righetti's thesis and spells out CBRN. It explains why the medium versus high risk distinction matters, and it states the three proxy-test critiques and the tool and interaction caveats accurately. It keeps his credit to the Preparedness Team and his pointed conclusion that OpenAI may not be able to tell either. A reader without specialist background can follow the actor, the development and the stakes. The question follows from the unfinished lab test the source describes.

I can't tell if o1-preview meets OpenAI's definition of "high risk." And I don't know if OpenAI can tell either.

date — pass: November 21, 2024 matches the originalPublished instant's UTC day and the source dateline. The evidence is a publisher page, so no archive disclosure is required. modifiedAt equals the publication time. The reading uses no later events or hindsight framing. It presents the wet lab study as still in development, as the source does.

Nov 21, 2024

support — pass: Each claim is source-stated and backed by its excerpt. The stated consequences of the risk levels match the source's description of OpenAI's voluntary commitments. The limitation about the funding disclosure is supported by the source's footnotes. No benefits, costs or forecasts are added beyond the source.

This benchmark was funded by my employer, Open Philanthropy, as part of our RFP on benchmarks for LLM agents.

media — pass: No media was selected and no verified media candidates were supplied, so nothing needs checking for identity or relevance.

Challenge or correct this reading

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