Examples
Examples from published documents
- OpenAI's incident plan sets procedures "for determining severity and notifying appropriate internal teams" once a potential incident is identified (OpenAI Frontier Governance Framework, §2.6)
- Anthropic rated a 2026 vendor-access incident as low risk: "Despite the features that make this incident notable, we assess the risk it posed to have been low." (Anthropic Risk Report, August 2026, §4.5.8.2.1)
Counter-examples
Not to be confused with
- Severity is not likelihood: a severity rating describes the harm of an incident that occurred or nearly occurred, while a risk level combines how bad an outcome would be with how likely it is.
- Severity is not incident type: the EU Code uses harm type to set deadlines, but two incidents of the same type (for example two cybersecurity breaches) can differ greatly in seriousness.
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.
- Intervention threshold
- Intervention threshold is OpenAI's term for a severity-based threshold that sets how it responds to misaligned model behaviour during training: immediate containment, mitigation during the training run, or mitigation before the next run.
- Source: OpenAI/Hugging Face Incident Technical Report
Source of record
Where this definition comes from
OpenAI Frontier Governance Framework, §2.6, Detection and triage
“"Once a potential AI safety incident is identified, the AIRP dictates procedures for assessing whether the event qualifies as an AI safety incident and for determining severity and notifying appropriate internal teams."”
https://cdn.openai.com/pdf/e37d949b-8c9f-4d76-b99e-4272f4631a7e/openai-frontier-governance-framework.pdfEU GPAI Code of Practice, Safety and Security Chapter, Commitment 9; Measure 9.2
“Signatories commit to "providing resourcing of such processes and measures appropriate for the severity of the serious incident and the degree of involvement of their model" (Commitment 9); "The level of detail in serious incident reports will be appropriate for the severity of the incident." (Measure 9.2)”
https://ec.europa.eu/newsroom/dae/redirection/document/118119OECD, Defining AI incidents and related terms (OECD Artificial Intelligence Papers No. 16, 2024), Serious AI incident, key clarifications
“"Serious AI incidents are a subset of AI incidents and are intended to include substantial, material incidents." "Assessing the seriousness of an AI incident is highly context-dependent; each jurisdiction may define it differently."”
https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/05/defining-ai-incidents-and-related-terms_88d089ec/d1a8d965-en.pdf
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| OpenAIShadow mapping OpenAI Frontier Governance Framework | “AIRP procedures "for determining severity and notifying appropriate internal teams" (§2.6).” The AIRP itself is not public, so the scale is unknown. | closeCL confidence: high | OpenAI Frontier Governance Framework |
| OpenAIShadow mapping OpenAI/Hugging Face Incident Technical Report | “"Enforcing intervention thresholds during reinforcement learning training runs to establish clear, severity-based thresholds for responding to misalignment, including thresholds for when OpenAI pursues immediate containment, mitigation during training, and mitigation before the next run." (§IX.C)” A planned measure; the thresholds are not published. | closeCL confidence: medium | OpenAI/Hugging Face Incident Technical Report |
| EUShadow mapping EU GPAI Code of Practice, Safety and Security Chapter | “Resourcing "appropriate for the severity of the serious incident and the degree of involvement of their model" (Commitment 9); report detail "appropriate for the severity of the incident" (Measure 9.2); deadlines tiered by harm type (Measure 9.3).” Severity is expressed through harm type and deadline tier, not a named scale. | closeCL confidence: high | EU GPAI Code of Practice, Safety and Security Chapter |
| EUShadow mapping EU AI Act (Regulation (EU) 2024/1689), Article 73(2) | “"The period for the reporting referred to in the first subparagraph shall take account of the severity of the serious incident." (Art. 73(2))” Applies to providers of high-risk AI systems, not to general-purpose AI model providers. | broadBR confidence: high | EU AI Act (Regulation (EU) 2024/1689), Article 73 |
| OECDShadow mapping OECD, Defining AI incidents and related terms (2024) | “Three nested tiers: "AI incident", "serious AI incident" ("a subset of AI incidents") and "AI disaster" ("a subset of serious AI incidents"); "Assessing the seriousness of an AI incident is highly context-dependent".” | closeCL confidence: high | OECD, Defining AI incidents and related terms (OECD Artificial Intelligence Papers No. 16, 2024) |
| Frontier Model ForumShadow mapping Frontier Model Forum -- Information Sharing Issue Brief | “Reporting regimes should address "the boundary between reportable incidents and lower-severity incidents or precursor events."” | broadBR confidence: medium | Frontier Model Forum -- Information Sharing Issue Brief |
| US Congress (S. 5061, 119th Congress, not enacted)Shadow mapping S. 5061, Sec. 4(a)(2)(D) | “NIST shall prioritise database inclusion of incidents that "would result in a high-severity or catastrophic impact to the people or economy of the United States" (Sec. 4(a)(2)(D)(ii)).” S. 5061 is a bill introduced in the US Senate on 21 July 2026; it has not been enacted. | broadBR confidence: medium | S. 5061 (119th Congress) |
| AnthropicShadow mapping Anthropic Risk Report (August 2026) | “Per-incident risk rating, e.g. "Despite the features that make this incident notable, we assess the risk it posed to have been low." (§4.5.8.2.1)” Uses the report's general risk-level words, not an incident-specific scale. | closeCL confidence: medium | Anthropic Risk Report (August 2026) |
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.
- UK AI Security Institute (AISI)
"We assessed each event for its potential to cause real-world harm." A per-event harm assessment with no published scale.
UK AI Security Institute -- Incident Report: unsanctioned agent behaviour during cyber testing (4 Aug 2026)
Divergence
Where sources materially disagree
Severity does different work in each source. OpenAI uses it to route internal escalation; the EU Code of Practice uses harm type to pick a 2-, 5-, 10- or 15-day clock and scales report detail to severity; the OECD and the Frontier Model Forum use it to separate reportable from lower-severity events; California SB 53 has no severity field because each of its four incident limbs already contains a harm threshold.
Gap
No source publishes a numbered incident-severity scale for frontier AI. OpenAI's AIRP severity levels and its RL intervention thresholds are described but not published, and the OECD leaves seriousness to each jurisdiction.
