Examples
Examples from published documents
- Anthropic runs "A Claude Opus 4.6 completion-only monitor covering completions with extended thinking by most internal users", and second-stage filtering surfaces cases that AI control researchers "manually review on the order of 50 trajectories per week" (Anthropic Risk Report, August 2026, §2.23.1.1)
- OpenAI lists as an example misalignment safeguard: "Monitor AI supervises agent actions to detect agent in the case of trying to take harmful actions, scheming, or similar" (OpenAI Preparedness Framework v2, Appendix C.2)
Counter-examples
Not to be confused with
- A blocking classifier that stops a response in real time is recorded as a classifier; a monitor may only flag activity for later review. Many systems use both.
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.
- Chain-of-thought monitoring
- Chain-of-thought monitoring is reading the step-by-step reasoning a model writes out, often with another model, to catch signs of intent to misbehave. Its proponents call it promising but imperfect, and fragile if development choices reduce how monitorable the reasoning is.
- Source: Korbak et al., Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety (arXiv 2507.11473)
- Post-market monitoring
- Post-market monitoring is the ongoing collection of information about a model after release, through user feedback, incident reports, bug bounties and usage analysis, to judge whether its systemic risks remain acceptable. The term is used in the EU GPAI Code of Practice.
- Source: EU GPAI Code of Practice, Safety and Security Chapter
Source of record
Where this definition comes from
Anthropic Risk Report, August 2026, §2.23.1.1
“A Claude Opus 4.6 completion-only monitor covering completions with extended thinking by most internal users, with no subsampling for traffic coming from the most common agent scaffolds, and a 5% subsampling for completions with extended thinking and one or more markers of using tools.”
https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdfOpenAI Preparedness Framework v2, Table 5
“Monitor AI supervises agent actions”
https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdfEU GPAI Code of Practice, Safety and Security Chapter, Measure 3.5
“will conduct appropriate post-market monitoring to gather information relevant to assessing whether the systemic risk could be determined to not be acceptable ... and to inform whether a Model Report update is necessary”
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 Risk Report, August 2026 / Training a Misaligned Reward Seeker | “"A Claude Opus 4.6 completion-only monitor covering completions with extended thinking by most internal users, with no subsampling for traffic coming from the most common agent scaffolds"; "A Claude Opus 4.8 prompt+completion monitor ... This covers a 25% subsample of prompts and completions in which a tool call might take an impactful action"; reviewers "manually review on the order of 50 trajectories per week" (§2.23.1.1); "we monitor the Chain-of-Thought to detect and measure reward hacking, but importantly do not reward or penalize the model based on it, in order to preserve monitorability"” Term also draws on Anthropic's "Training a Misaligned Reward Seeker" for the CoT-monitoring/severity-flagging quote. | exactEQ confidence: high | Anthropic Risk Report, August 2026 |
| OpenAIShadow mapping OpenAI Preparedness Framework v2 / Pacing Model Development (Cyber), Aug 2026 | “"Monitor AI supervises agent actions to detect agent in the case of trying to take harmful actions, scheming, or similar" (PF v2, App. C.2); OpenAI's Pacing Model Development post (Aug 2026) describes monitoring as a safeguard "which detects and allows us to respond to concerning behavior", with activation classifiers, "detectors that inspect a model's internal activity", escalating to "increasingly sophisticated and high-compute automated investigators"” Only the first quote is from PF v2. The other quotes are from OpenAI's Aug 2026 post, read through a reader proxy because openai.com refused automated requests; CASRAI has not verified them against the live page. | exactEQ confidence: high | OpenAI Preparedness Framework v2 |
| Google DeepMindShadow mapping Frontier Safety Framework v3.1 | “Deployment mitigations include "input/output/chain-of-thought monitoring and analysis" (glossary); safeguards "such as chain-of-thought monitoring on high-risk deployments"” | closeCL confidence: medium | Google DeepMind Frontier Safety Framework v3.1 |
| xAIShadow mapping Frontier AI Framework, 30 Jun 2026 | “Incident detection channel "Monitoring and alerting of public comments from the X platform"; security measures include real-time monitoring (described, not quoted)” xAI's Frontier AI Framework (30 Jun 2026) carries PDF metadata reading "Privileged/Confidential DRAFT working FRAMEWORK DOC" with no xAI statement found disambiguating draft from final; the cited URL also returned HTTP 404 (NoSuchKey) on 13 Nov 2026, so the quote is verified against the local capture only. Treat as provisional. | narrowNR confidence: medium | xAI Frontier AI Framework, 30 June 2026 (draft-labeled PDF metadata) |
| MetaShadow mapping Meta Advanced AI Scaling Framework v2 | “LlamaFirewall "chain-of-thought auditing" and input classifiers” | narrowNR confidence: medium | Meta Advanced AI Scaling Framework v2 |
| EUShadow mapping EU GPAI Code of Practice, Safety and Security Chapter | “Measure 3.5 "Post-market monitoring": Signatories "will conduct appropriate post-market monitoring to gather information relevant to assessing whether the systemic risk could be determined to not be acceptable ... and to inform whether a Model Report update is necessary", via methods including end-user feedback, incident reporting forms, bug bounties, "monitoring software repositories, known malware, public forums, and/or social media", and "monitoring aspects of models that are relevant for assessing and mitigating systemic risk and are not transparent to third parties, e.g. hidden chains-of-thought"” This confirms xAI's chapter-derived s.2.2(5) wording; xAI's framework carries the draft/final metadata caveat above. | exactEQ confidence: high | EU GPAI Code of Practice, Safety and Security Chapter |
| Frontier Model ForumShadow mapping FMF Information Sharing / Incident Reporting Issue Brief | “"Monitoring and Detection Systems: Enhancing systems that detect anomalous behavior, unauthorized access, or potential misuse, for safety and security purposes only" (Table 3)” | closeCL confidence: medium | Frontier Model Forum, Information Sharing / Incident Reporting Issue Brief |
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.
- SB 53 (California)
Framework topic "(10) ... including risks resulting from a frontier model circumventing oversight mechanisms" (22757.12(a)) — related to monitoring but a pointer, not a mapping.
California SB 53 - UK AISI / Google DeepMind
"CoT monitoring helps us understand how an AI system produces its answers, complementing interpretability research." — related, not a direct monitor-record mapping.
DeepMind, Deepening Our Partnership with UK AI Security Institute







