Examples
Examples from published documents
- Anthropic's August 2026 Risk Report states: "This report is not scoped to a single AI model. Rather, it is a risk assessment of Anthropic's activities as a whole" (Anthropic Risk Report, August 2026, §1)
- Anthropic's Responsible Scaling Policy commits: "We will publish a Risk Report every 3-6 months", each covering risks "as of a given coverage date" (Anthropic Responsible Scaling Policy v3.4, §3.1)
- An EU Model Report must give "a detailed justification for why the systemic risks stemming from the model are acceptable, including details of the safety margins incorporated" (EU GPAI Code of Practice, Safety and Security Chapter, Measure 7.2)
Counter-examples
Not to be confused with
- A safety framework is not a risk report: the framework states how risks will be assessed and handled, while a risk report applies it and gives the result as of a date.
- A model card or system card describes one model's capabilities and testing; Anthropic's Advanced AI Framework treats system cards as per-model updates to the risk report's assessment, not as replacements for it.
Related terms
Related AI safety terms
Terms defined under this element that are not NIKOLAI elements themselves, each as the cited document uses it. Every term is also listed in the AI safety glossary.
- Capabilities Report
- A Capabilities Report is OpenAI's internal compilation of a covered model's evaluation results and noteworthy observations, submitted to its Safety Advisory Group to decide whether a capability threshold has been reached.
- Source: OpenAI Preparedness Framework v2
- Safety and Security Model Report
- A Safety and Security Model Report ("Model Report") is the document an EU GPAI Code signatory gives the AI Office before placing a model on the market, describing the model, its systemic risk assessment and mitigations, and why the risks are acceptable.
- Source: EU GPAI Code of Practice, Safety and Security Chapter
- Transparency report (SB 53)
- A transparency report, under California SB 53, is published before or with a new frontier model's deployment; large developers must summarise their catastrophic risk assessments, the results, and how far third-party evaluators were involved.
- Source: California SB 53
Source of record
Where this definition comes from
Anthropic Advanced AI Framework, p.6
“"Risk reports. Require Covered Developers to publish a risk report at least every six months that offers an overall assessment of each of the Enumerated Risks posed by the developer's models." Contents include "An assessment of the risk in each Enumerated Risk category that remains after accounting for the safety mitigations the developer has in place." (p.6)”
https://www-cdn.anthropic.com/files/4zrzovbb/website/0a58d567024a8b448ff15158ebc3625328dfcc1f.pdfEU GPAI Code of Practice, Safety and Security Chapter, Commitment 7
“"Signatories commit to reporting to the AI Office information about their model and their systemic risk assessment and mitigation processes and measures by creating a Safety and Security Model Report ("Model Report") before placing a model on the market" (Commitment 7)”
https://ec.europa.eu/newsroom/dae/redirection/document/118119
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| AnthropicShadow mapping Anthropic Responsible Scaling Policy v3.4 | “"A Risk Report will cover the risks of Anthropic's models and actions as of a given coverage date, which will be no more than 30 days prior to the publication of the report." "We will publish a Risk Report every 3-6 months." (§3.1)” Verified against CASRAI's local capture of the RSP v3.4 PDF (extracted 16 Sep 2026). | exactEQ confidence: high | Anthropic Responsible Scaling Policy v3.4 |
| AnthropicShadow mapping Anthropic Risk Report, August 2026 | “"This report is not scoped to a single AI model. Rather, it is a risk assessment of Anthropic's activities as a whole. We consider all of our models, including those we run only internally, in our assessment." (§1)” | exactEQ confidence: high | Anthropic Risk Report, August 2026 |
| AnthropicShadow mapping Anthropic Advanced AI Framework | “Recommends that Covered Developers "publish a risk report at least every six months that offers an overall assessment of each of the Enumerated Risks posed by the developer's models", with capabilities, "The threat models being tracked, how observed capabilities map to them, and the mitigations addressing each", and residual risk (p.6).” A legislative proposal by Anthropic, not an enacted requirement. | exactEQ confidence: high | Anthropic Advanced AI Framework |
| EUShadow mapping EU GPAI Code of Practice, Safety and Security Chapter | “"Safety and Security Model Report ("Model Report")" created "before placing a model on the market" and kept up to date (Commitment 7); it must give "a detailed justification for why the systemic risks stemming from the model are acceptable, including details of the safety margins incorporated" (Measure 7.2).” Per model rather than per company, and addressed to the EU AI Office rather than published. | closeCL confidence: high | EU GPAI Code of Practice, Safety and Security Chapter |
| OpenAIShadow mapping OpenAI Preparedness Framework v2 | “Two per-model internal reports: evaluation results "are compiled into a Capabilities Report that is submitted to the SAG" (§3.3); "We compile the information on the planned safeguards needed to minimize the risk of severe harm into a Safeguards Report", which includes "An assessment on the residual risk of severe harm based on the deployment" (§4.2).” Narrower: each report covers one part (capabilities or safeguards) of one model's assessment and is addressed to OpenAI's Safety Advisory Group. | narrowNR confidence: high | OpenAI Preparedness Framework v2 |
| MetaShadow mapping Meta Advanced AI Scaling Framework v2 | “Meta incorporates evaluation results "into our preparedness reports", and "Preparedness reports will also disclose any known issues that could hinder generalizing our safety testing to real-world risks" (§2.1, §2.2.1).” | closeCL confidence: medium | Meta Advanced AI Scaling Framework v2 |
| Google DeepMindShadow mapping Gemini 3.7 Flash FSF Report | “Per-model "Frontier Safety Framework Report – Gemini 3.7 Flash" (title page), reporting for each risk domain whether a T/CCL was reached and the resulting deployment determination.” A per-model report under the framework; the report does not describe itself as a risk report (CASRAI reading). | narrowNR confidence: medium | Gemini 3.7 Flash FSF Report |
| California SB 53Shadow mapping California SB 53 | “Transparency report published "Before, or concurrently with, deploying a new frontier model"; a large frontier developer must include summaries of "(A) Assessments of catastrophic risks from the frontier model conducted pursuant to the large frontier developer's frontier AI framework. (B) The results of those assessments. (C) The extent to which third-party evaluators were involved." (22757.12(c)(1)-(2))” Verified against CASRAI's local capture of the statute text; leginfo.legislature.ca.gov blocked automated retrieval on 2026-11-13. Statute does not define the term "transparency report". | narrowNR confidence: high | California SB 53 |
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.
- METR
A risk report written by a third-party evaluator about several developers, not a developer's own risk report; a pointer, not a mapping.
METR Frontier Risk Report (pilot, Feb-Mar 2026)
Divergence
Where sources materially disagree
Sources disagree on three axes. Scope: Anthropic's Risk Report covers "Anthropic's activities as a whole", while the EU Model Report, OpenAI's Capabilities and Safeguards Reports and Google DeepMind's FSF reports are per model. Audience: Anthropic, Google DeepMind and SB 53 transparency reports are public; the EU Model Report goes to the AI Office; OpenAI's reports go to its internal Safety Advisory Group. Cadence: every 3-6 months (Anthropic RSP v3.4), every six months (Anthropic's AAF proposal), before market placement with updates (EU), before or with each deployment (SB 53).
Gap
CASRAI did not find a shared section order or machine-readable format for risk reports in these sources; SB 53 lists required contents but sets no format.
