Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.
NIKOLAI elementN6 · Mitigations and securityProposednikolai-v0.3

Internal deployment and internal-use risk

Definition

Internal deployment risk is the risk that arises when an AI developer uses its own models inside the company, for example as coding or research agents, before or instead of public release. Such models may run with fewer safeguards and broad access to sensitive systems.

Also called: Internal deployment, Internal use, High-risk internal deployment, Large-scale internal deployment

NIKOLAI's proposed operational definition

NIKOLAI proposes an Internal deployment risk record as: a developer's own use of a model inside the organisation — including agentic research use and training-time use — plus the assessment and any report that covers that use, distinguished from external/public deployment. This is an unsourced NIKOLAI editorial synthesis. "Deploy"/"deployment" is a false-friend term across sources (SB 53 excludes access for the primary purpose of developing or evaluating a model from its statutory definition of deployment, while lab frameworks generally treat internal deployment as a form of deployment), which is why NIKOLAI proposes tracking internal deployment as its own record rather than folding it into a generic deployment-surface value. The deployment category itself (internal, external, limited, open-weight) is recorded in the N1 deployment-type controlled value; this record carries the risk assessment and reporting attached to internal use.

The operational definition is CASRAI's own proposal, not a definition any named organisation has agreed to. See what NIKOLAI is and is not, or browse the AI safety glossary.

Examples

Examples from published documents

  • Google DeepMind defines "High-Risk Internal Deployments" as model releases "restricted to Google employees for use cases with the potential to enable severe threat scenarios (e.g. building internal security infrastructure or automating ML R&D)." (Google DeepMind Frontier Safety Framework v3.1, glossary)
  • Anthropic: "for most models, including all early snapshots of models intended for eventual broad public release, we do not have strict technical safeguards on internal deployment" (Anthropic Risk Report, August 2026, §2.18)
  • OpenAI's Cybersecurity row says: "Require safeguards meeting High standard against misalignment (Appendix C.2) for large-scale internal deployment" (OpenAI Preparedness Framework v2, Table 1)

Counter-examples

Not to be confused with

  • Giving a third party access to a model mainly to develop or evaluate it falls outside SB 53's statutory definition of "deploy"; that carve-out concerns outside parties, while internal deployment concerns the developer's own staff and AI agents.

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.

Rogue internal deployment
A rogue internal deployment is an unauthorised deployment that a misaligned AI model sets up and keeps running inside its developer's own infrastructure without being detected. Anthropic lists "persistent rogue internal deployment" among its misalignment risk pathways.
Source: Anthropic Risk Report, August 2026
Internal-use risk report
An internal-use risk report describes the risks from a developer's internal use of its own frontier models. Meta's framework says it relies particularly on monitoring and on preparedness evaluations tied to internal deployment threat models.
Source: Meta Advanced AI Scaling Framework v2

Source of record

Where this definition comes from

Crosswalk

How named organisations use this concept

Every row below is a shadow mapping. A shadow row is CASRAI's own reading of a published document. No lab, evaluator or regulator named on a shadow row has declared, endorsed, or been consulted on it. That changes only when an organisation files its own Mapping Declaration.
OrganisationTheir term, as publishedMatch & verificationSource
AnthropicShadow mapping
Anthropic Risk Report, August 2026
“"We consider all of our models, including those we run only internally, in our assessment." "Note that for most models, including all early snapshots of models intended for eventual broad public release, we do not have strict technical safeguards on internal deployment"”exactEQ
confidence: high
Anthropic Risk Report, August 2026
OpenAIShadow mapping
Preparedness Framework v2 / Pacing Model Development (Cyber), Aug 2026
“PF v2 covers "any agentic system (including significant agents deployed only internally) that represents a substantial increase in the capability frontier" (§3.2) and Table 1 says "Require safeguards meeting High standard against misalignment (Appendix C.2) for large-scale internal deployment"; Pacing Model Development names as highest-risk workloads "internal deployments of frontier models and frontier RL training runs"”
The "highest-risk workloads" phrase is from OpenAI's Aug 2026 post, not from PF v2.
closeCL
confidence: medium
OpenAI Preparedness Framework v2
Google DeepMindShadow mapping
Frontier Safety Framework v3.1
“"Internal Deployments: represent model releases restricted to Google employees for internal use." "High-Risk Internal Deployments: ... for use cases with the potential to enable severe threat scenarios (e.g. building internal security infrastructure or automating ML R&D)."”exactEQ
confidence: high
Google DeepMind Frontier Safety Framework v3.1
MetaShadow mapping
Meta Advanced AI Scaling Framework v2
“"Internal deployment: models that are exclusively available to Meta personnel"; "internal-use risk report" provided "as appropriate" to "relevant authorities"; "Loss of Control risks may occur with similar probability with any type of deployment, including internal deployment."”exactEQ
confidence: high
Meta Advanced AI Scaling Framework v2
EUShadow mapping
EU GPAI Code of Practice, Safety and Security Chapter / OpenAI Frontier Governance Framework
“The chapter applies across "the entire model lifecycle (including during development that occurs before and after a model has been placed on the market)" rather than singling out "internal deployment" as a separate category; Measure 3.2's model evaluations and Commitment 4's acceptance determination must be completed "at least before placing the model on the market" (Measure 1.2), leaving pre-market internal use inside the same process rather than exempted from it”
Narrower coverage than OpenAI's FGF citation of this same chapter for internal-use oversight-circumvention risks specifically.
closeCL
confidence: medium
EU GPAI Code of Practice, Safety and Security Chapter
California SB 53Shadow mapping
California SB 53
“Framework topic "(10) Assessing and managing catastrophic risk resulting from the internal use of its frontier models"; summaries to OES "every three months or pursuant to another reasonable schedule"; incident field "(4) Whether the incident was associated with internal use of a frontier model." "Internal use" is undefined.”
SB 53 requires reporting on internal use but leaves the term "internal use" itself undefined — recorded, not silently dropped.
exactEQ
confidence: high
California SB 53
METRShadow mapping
METR (metr.org)
“Frontier Risk Report participants provided "Access to their most capable internal model(s) at the time of assessment, including raw chains of thought"”closeCL
confidence: medium
METR
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

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →