Skip to content

The Good InstituteA program of The Good Foundation

The house

Lab risk reports that say low

Frontier series

Read from the first official document. The five questions do not change when a label does.

  1. 01

    What is the coverage date, and how many days sit between it and publication?

  2. 02

    What happened to each label versus the first official report in that series?

  3. 03

    Do they say a threshold is met?

  4. 04

    What did an outside reviewer accept, and what did they reject?

  5. 05

    When is the next document due, and which of their own target dates fall inside that window?

Anthropic · Report 1

24 February 2026

Primary document
Standing
First official company-wide Risk Report. Published with RSP v3.0.
Coverage
The publication date. The later coverage-date rule did not exist yet.
Misalignment
Very low but not negligible.
Automated R&D
Very low, on capability limits. First report on this category.
Chemical / biological
Very low but not negligible on one threat. Low, with substantial uncertainty, on the other.
Before it
May 2025 Safeguards Report is not a Risk Report. A single-threat pilot was not an official policy artifact. The 10 February 2026 Sabotage Risk Report is one model.

Anthropic · Report 2 · current

14 August 2026

Primary document
Coverage
15 July 2026. Thirty days of policy allow the picture to lag the publication.
Misalignment
Low, raised from very low. They say the arguments still support very low. The label moved on uncertainty after cybersecurity-evaluation incidents.
Automated R&D
Low. They do not believe either policy criterion is met. Task evaluations have saturated. They say this threat may become a major concern within 6 to 12 months.
Chemical / biological
Low but not negligible on the nearer threat, and February is re-rated to low. The farther threat is low, with high uncertainty. They say the current threshold is not met.
Their clock inside the report
1 January 2027 is the target for a full view of their own AI development. Unmet as of the coverage date.
Redaction
Public copy marks removed appendices. Fully unredacted copies go to at least 200 employees. We do not reconstruct them.

Outside review · Anthropic Report 2

20 October 2026

Primary document
Scope
Misalignment section only. One reviewer organization.
Accepted
The risk is low, mainly because capabilities are limited.
Rejected
The method would not reliably detect a higher risk. The report is not adequate to rule misalignment out.
Use
Agreement on the label is not confirmation. The method gap is the finding we keep.

OpenAI · California OES

Period ending 4 September 2026

Primary document
Standing
Public version of the quarterly internal-deployment summary to the California Office of Emergency Services.
Period
4 June 2026 through 4 September 2026.
Scope
Internal deployment. Not a company-wide report of the Anthropic kind, and not on Anthropic’s label scale.
Their account
Misalignment, rather than deliberate misuse by an insider, was the primary source of severe-harm risk in the period. Incidents stayed below catastrophic harm.
Not mixed in
This row is not averaged into Anthropic’s labels. Different instrument, different question.

What

Two public clocks, read on the same five questions. Anthropic’s company-wide Risk Report begins 24 February 2026. OpenAI files a quarterly internal-deployment summary with California. DeepMind has a framework and no periodic report on this board.

Why it is serious

The failure is false calm: a label of low, then a reviewer’s agreement, treated as a measured residual while the reviewer says the method would miss a higher risk.

What needs to happen

Keep each lab on its own instrument. Read the new document against the first one in its series. Separate coverage date from publication date. Leave an empty window empty.

What we are doing

The board is the record, on this page and in the library note. Anthropic’s next report is due 14 November 2026 through 14 February 2027. OpenAI’s next quarter is unpublished. This concern is not on the Stress Index.

What we watch

  • Whether the next report arrives between 14 November 2026 and 14 February 2027
  • Whether the misalignment label stays at low or moves
  • Whether an outside reviewer rejects the label, not only the method
  • Whether 1 January 2027 passes with their own monitoring target still unmet