Examples
Examples from published documents
- Anthropic's RSP v3.3 "revises our threshold for novel chemical/biological weapons production to better track the threat model of concern" (Anthropic RSP v3.4, changelog)
- Meta changed the consequence of its Critical threshold from "Stop" to "Develop with Mitigations" in version 2 (Meta Advanced AI Scaling Framework v2, Appendix II)
- "Under the updated FSF, which is more appropriately calibrated to real world harm, the alert threshold for CBRN has not been reached for Gemini 3 Pro" (Gemini 3 Pro FSF report, p.4)
Counter-examples
Not to be confused with
- A change to evaluation cadence or to required safeguards is a framework change, but not a threshold change unless the threshold's own wording, level or consequence changes.
Source of record
Where this definition comes from
Anthropic Responsible Scaling Policy v3.4, Changelog, pp.20-21
“Changelog, RSP v3.4: "It revises the Automated R&D capability threshold in light of further issues that have come up in discussion". Changelog, RSP v3.3: "This update (1) revises our threshold for novel chemical/biological weapons production to better track the threat model of concern".”
https://cdn.sanity.io/files/4zrzovbb/website/0bacdc8440ea96e62a8766d99ebe1d4eea6d5f3a.pdfMeta Advanced AI Scaling Framework v2, Appendix II, p.44
“"Critical threshold changed from 'Stop' to 'Develop with Mitigations.' High threshold measure changed from 'Do not release' to 'Deploy with mitigations.' Moderate or lower threshold measure changed from 'Release' to 'Deploy.'"”
https://ai.meta.com/static-resource/Meta_Advanced-AI-Scaling-Framework-v2
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| AnthropicShadow mapping Responsible Scaling Policy v3.4 | “Changelog, v3.4: "It revises the Automated R&D capability threshold in light of further issues that have come up in discussion"; v3.3: "revises our threshold for novel chemical/biological weapons production to better track the threat model of concern"; v3.1 addresses "how we operationalize the Automated R&D capability threshold". Policy-change process: "we will publicly share the updated version prior to or on its effective date and will record the differences from the prior draft in the Change Log." (§4(8))” | exactEQ confidence: high | Anthropic Responsible Scaling Policy v3.4 |
| MetaShadow mapping Meta Advanced AI Scaling Framework v2 | “"Critical threshold changed from 'Stop' to 'Develop with Mitigations.' High threshold measure changed from 'Do not release' to 'Deploy with mitigations.' Moderate or lower threshold measure changed from 'Release' to 'Deploy.'" (Appendix II, p.44)” Records the change to each threshold's consequence, not to the threshold wording. | exactEQ confidence: high | Meta Advanced AI Scaling Framework v2 |
| MagicShadow mapping Magic AGI Readiness Policy | “"we may update this threshold upward over time. We may also modify the public and private benchmarks used. Such a change will require approval by our Board of Directors, with input from external security and AI safety advisers." (item 3)” Names the approver of a threshold change. The page carries the banner "Note: This policy is outdated. We're working on an update." | exactEQ confidence: high | Magic AGI Readiness Policy |
| OpenAIShadow mapping Preparedness Framework v2 | “"we are removing terms “low” and “medium” from the Framework, because those levels were not operationally involved in the execution of our Preparedness work." (summary of changes from v1)” | closeCL confidence: high | OpenAI Preparedness Framework v2 |
| Google DeepMindShadow mapping Frontier Safety Framework v3.1 | “After a framework assessment Google may "Update our risk domains and T/CCLs, where necessary." (s.5.1). A response plan "may include updating the alert threshold to ensure the safety buffer remains appropriate." (glossary, Response Plans)” | closeCL confidence: high | Google DeepMind Frontier Safety Framework v3.1 |
| G42Shadow mapping G42 Frontier Safety Framework | “Threshold table action once a threshold is reached: "Update the policy to include more advanced thresholds" (s.2).” | closeCL confidence: medium | G42 Frontier Safety Framework |
| AI Seoul Summit (Frontier AI Safety Commitments)Shadow mapping Frontier AI Safety Commitments | “Approaches "may evolve in the future. In such instances, organisations will provide transparency on this, including their reasons, through public updates." (preamble, P2)” Covers changes to any part of a signatory's approach, not thresholds specifically. | broadBR confidence: high | Frontier AI Safety Commitments (AI Seoul Summit 2024) |
What do these codes mean?
- exact
- The source term is equivalent to this element
- close
- The source term is close but not equivalent to this element
- broad
- The source term is broader than this element
- narrow
- The source term is narrower than this element
- none
- No mapping claim — used for false-friend and declared-but-undefined rows
- EQ
- Equivalent
- CL
- Close
- BR
- Source is broader than the element
- NR
- Source is narrower than the element
- FF
- False friend — same or similar label, different meaning
- DU
- Declared but undefined by the source
- UV
- Unverified
Related, not mapped
Pointers that are not crosswalk claims
These sources mention this concept but do not define or map it clearly enough to count as a crosswalk row — noted here so the research is visible without overstating it as a mapping.
- Frontier Model Forum
"Static Historical Standards": "risk thresholds would be established at a specific point in time (e.g. 2023) and never adjusted, regardless of changes in the ecosystem" (s3.2). The report presents this as one of two baselining approaches, the alternative being marginal risk assessment.
FMF Risk Taxonomy and Thresholds
Divergence
Where sources materially disagree
Sources record different things under a threshold change. Anthropic records threshold-wording changes per point release with a one-line rationale; Meta records changes to the consequence attached to each threshold ("Stop" to "Develop with Mitigations"); OpenAI removed whole levels; Google DeepMind allows alert thresholds to move inside a response plan without a framework version change; Magic requires Board of Directors approval to raise a threshold or change a benchmark, and Anthropic's RSP changes are proposed by the CEO and Responsible Scaling Officer and approved by the Board in consultation with the Long-Term Benefit Trust.
Gap
Anthropic's Risk Report gives the replaced AI R&D wording ("compress two years of 2018-2024 AI progress into a single year") next to the current test, but most changelogs give a summary without the old and new threshold text side by side. Google DeepMind's Gemini 3 Pro report notes that its CBRN alert-threshold result changed under the "more appropriately calibrated" FSF v3 without publishing the old and new thresholds.
