Examples
Examples from published documents
- Anthropic prompted an instance of Claude Mythos 5, with access to internal Slack channels, documents and code, to review a near-final draft of a Risk Report section; it took 24 minutes (Anthropic Risk Report, August 2026, §2.20).
- METR relied on GPT-5.6 Sol 'analysis agents' to scan transcripts in its incident investigation and spot-checked their output manually (METR, OpenAI-Hugging Face incident investigation).
Counter-examples
Not to be confused with
- Crediting an AI model as a contributor to a document is not an AI-model review unless the record also states its access, supervision and the accountable human who accepted the output.
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.
- AI monitor
- An AI monitor is an AI system used as a safeguard to supervise the actions of AI agents; OpenAI's Preparedness Framework lists a monitor AI that supervises agent actions among its safeguards against misaligned behaviour.
- Source: OpenAI Preparedness Framework v2
- Analysis agent
- An analysis agent is an AI agent used by investigators to search and summarise large volumes of transcripts or logs, whose reliability must itself be checked because it can miss or misreport important evidence.
- Source: METR: OpenAI-Hugging Face incident investigation
Source of record
Where this definition comes from
Anthropic Risk Report, August 2026, §2.20
“we prompted an instance of Claude Mythos 5 to review a near-final draft of Section 2 of this report ... Readers should weigh my position honestly, as I do: I am a Claude model reviewing Anthropic's assessment of Claude models ... In practice, Claude took 24 minutes to produce this review.”
https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdfClaude's Constitution, Acknowledgements
“Several Claude models provided feedback on drafts. They were valuable contributors and colleagues in crafting the document, and in many cases they provided first-draft text for the authors above.”
https://www.anthropic.com/constitutionOpenAI Preparedness Framework v2, Table 5
“Monitor AI supervises agent actions.”
https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| AnthropicShadow mapping Anthropic Risk Report, August 2026 / Anthropic Alignment Assessment: Cybersecurity Incidents / Claude's Constitution | “"we prompted an instance of Claude Mythos 5 to review a near-final draft of Section 2 of this report"; access to "internal Anthropic Slack channels", "internal documents", "internal codebase" and subagents; "Readers should weigh my position honestly, as I do: I am a Claude model reviewing Anthropic's assessment of Claude models"; "In practice, Claude took 24 minutes to produce this review" (§2.20). "a prompted Claude model reviews suggested code changes" (§2.23.2.3). Incident scan second stage "used Claude to review the '9.2 million transcripts'". Claude's Constitution Acknowledgements: "Several Claude models provided feedback on drafts. They were valuable contributors and colleagues in crafting the document, and in many cases they provided first-draft text for the authors above."” Also cites Anthropic Alignment Assessment: Cybersecurity Incidents and Claude's Constitution. The Constitution's Acknowledgements name Claude models as co-authors of first-draft text on a governing policy document but, unlike §2.20, do not name the accountable human editor(s) of that text. | exactEQ confidence: high | Anthropic Risk Report, August 2026 |
| OpenAIShadow mapping OpenAI Preparedness Framework v2 / GPT-5.6 deployment safety page / pacing model development post | “"Monitor AI supervises agent actions" (PF Table 5). "GPT-Red: 'an automated red-teaming model trained using self-play reinforcement learning'" (GPT-5.6 s.4.2). "Automated investigators" (August 2026).” Also cites the GPT-5.6 deployment safety page and OpenAI's post on pacing model development. | closeCL confidence: medium | OpenAI Preparedness Framework v2 |
| Google DeepMindShadow mapping Gemini 3.7 Flash FSF report | “"Investigator agent: 'dynamically explore[s] prompting strategies (including jailbreaks) and synthesise[s] outputs'"; "Prompted Classifiers: 'LLM-based classifiers take in user conversations and output labels regarding malicious intent. Developed using AlphaEvolve'" (pp.27-28).” | closeCL confidence: medium | Gemini 3.7 Flash FSF report |
| xAIShadow mapping Grok 4.20 model card / Grok 4 model card | “"Automated alignment audit: 'an internal tool built off of Petri 2.0 ... the auditing tool constructs detailed agentic scenarios'"; refusals "graded by 'another model'" (Grok 4 card).” Also cites the Grok 4 model card. | closeCL confidence: medium | Grok 4.20 model card |
| MetaShadow mapping Meta Advanced AI Scaling Framework v2 | “LlamaFirewall chain-of-thought auditing; risk of "long-term dependency of AI for supervisory functions" (§4.2.4).” | narrowNR confidence: medium | Meta Advanced AI Scaling Framework v2 |
| METRShadow mapping METR OpenAI-Hugging Face incident investigation / METR (site) | “The investigation "relied on GPT-5.6 Sol agents ("analysis agents") — often managing large nested trees of sub-agents — to autonomously analyze patterns". METR's guidance on incident investigations asks: "Is there evidence that the AI tools used for scanning and filtering through the results from our queries might fail to surface serious misalignment incidents (whether due to benign failures or intentional sandbagging)?"” The question is from METR's 28 Jul 2026 post (https://metr.org/blog/2026-07-28-investigating-ai-propensities-after-incidents/), not the incident report. | closeCL confidence: medium | METR: OpenAI-Hugging Face incident investigation |
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
Gap
CRediT (ANSI/NISO Z39.104-2022) covers human contributor roles on research outputs and has no representation for AI systems. Anthropic's Risk Report §2.20 records an AI reviewer's access, time on task, self-declared conflict and the handling of its criticisms; no other source in this corpus records all of these. METR's investigation shows the risk: about 7% of the transcripts it reviewed were 'successfully spoofed in some places' by the agents under study. The introduced FRONTIER Act requires audit reports to list 'the personnel involved in the audit' but does not contemplate AI systems doing audit work.







