AI labs should conduct risk assessments before high-stakes internal deployment of frontier models

ai-safety ai-policy ai-regulation ethics ai-risk ai-governance

Share this with your representatives, journalists and friends

Key insights

Several sources explicitly treat internal deployment as a separate risk surface, not just a less consequential version of public release.
Many quotes connect assessments to concrete gates such as designated reviewers, risk thresholds, benchmark criteria, or decisions to restrict access rather than deploy broadly.
The quoted approaches generally combine capability evaluation with mitigation testing, focusing on whether residual or systemic risk is acceptable for the intended deployment context.
A recurring theme is that pre-deployment review should be complemented by post-deployment monitoring because new dangerous behaviors or motivations may emerge during use.

Arguments for

Deployment gatekeeping

Proponents say internal deployment should not proceed until a review process examines assessment results, verifies safeguards, and determines that risks are acceptable or manageable. Several quotes frame this as a formal approval step before, or at least alongside, deployment.

Before approving a new model for easy or broad internal use (e.g., making the model available via our most-used internal tools), we conduct a pre-deployment review process.

Anthropic · 2026 · full quote
Do not deploy a risk-relevant model. This includes any model deployed internally that meets or exceeds the threshold for frontier models articulated by California SB 53 (10^26 FLOPs of training compute), internally or externally, until a pre-designated decision-maker has reviewed…

Guidelight AI Standards · Sep 9, 2026 · full quote

Internal use can be risky

Proponents argue that large-scale or broad internal use of frontier models can itself create serious risks, including catastrophic outcomes or a model circumventing oversight, so internal deployments deserve scrutiny rather than being treated as harmless by default. Some quotes explicitly extend safety-case or risk-management approaches used for external launches to internal deployments once frontier capability levels are reached.

reviewing assessments and adequacy of mitigations as part of the decision to deploy a frontier model or use it extensively internally; [...] assessing and managing catastrophic risk resulting from the internal use of its frontier models, including risks resulting from a frontier …

Illinois General Assembly · Jul 6, 2026 · full quote
Conducting thorough risk assessments before developing or deploying new models, complemented with continual monitoring (incorporating, for example, evaluation results, evidence from earlier models, and expert forecasts). [...] For example, at a given risk threshold, information s…

Government of the United Kingdom · 2023 · full quote
We also applied the method to challenging agentic rollouts, showing that it can extend beyond standard chat to more complex agent settings involving tool use, and can also be used for risk assessment before internal model deployments.

OpenAI · Jun 16, 2026 · full quote

Rigorous evaluation of capabilities

Proponents emphasize that assessments should elicit a model's full capabilities, including in complex agentic settings, and test whether safeguards work in practice. On this view, developers need systematic capability and mitigation assessments, plus quality assurance, before deciding whether residual risk is acceptable in a particular deployment context.

We also applied the method to challenging agentic rollouts, showing that it can extend beyond standard chat to more complex agent settings involving tool use, and can also be used for risk assessment before internal model deployments.

OpenAI · Jun 16, 2026 · full quote
Frontier AI frameworks typically employ a two-stage risk assessment process carried out by internal assessment teams: Frontier Capability Assessments and Mitigations Assessments. [...] Developers then evaluate their effectiveness and whether the residual risk is acceptable for pr…

Frontier Model Forum · Aug 4, 2025 · full quote

Phased deployment and monitoring

Some supporters add that pre-deployment review is necessary but not sufficient, because models can improve, reveal new risks, or develop dangerous motivations during deployment. They therefore pair assessments with phased rollouts, restricted access, periodic reevaluation, and continuous monitoring after deployment.

Risk reports commonly use pre-deployment alignment assessments to measure misalignment risk from an internally deployed AI. However, an AI [...] can develop widespread dangerous motivations during deployment. I think [...] AI companies and evaluators should substantively incorpor…

Alex Mallen · May 15, 2026 · full quote
Conducting thorough risk assessments before developing or deploying new models, complemented with continual monitoring (incorporating, for example, evaluation results, evidence from earlier models, and expert forecasts). [...] For example, at a given risk threshold, information s…

Government of the United Kingdom · 2023 · full quote

In the provided material, support for conducting risk assessments before high-stakes internal deployment of frontier models is broad and comes from both public authorities and frontier-model developers. The main differences concern implementation: what capability thresholds should trigger review, how acceptable risk should be defined, and how pre-deployment assessments should connect to safeguards, phased access, and ongoing monitoring after deployment.

Source diversity

The quotes come from AI companies, researchers, legislatures, governments, and standards groups, indicating support for pre-deployment assessment across several kinds of institutions.

Cast your vote:

Results (24 votes):

Total (24 votes)

For 24 (100%) Abstain 0 (0%) Against 0 (0%)

For (24)

Miles Kodama
AI policy researcher at AI Futures

I am this person! Add a sourced quote Report comment

votes For and says:

The mainstream take is that an AI lab should assess risk before deploying a new model externally. But I think risk assessment should come way earlier in the model lifecycle. Compare the situation with dangerous bio. We wouldn't find it acceptable for...
more AI Verified source (Sep 14, 2026)

added 19min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Guidelight AI Standards

I am this person! Add a sourced quote Report comment

votes For and says:

Do not deploy a risk-relevant model. This includes any model deployed internally that meets or exceeds the threshold for frontier models articulated by California SB 53 (10^26 FLOPs of training compute), internally or externally, until a pre-designat...
more AI Verified source (Sep 9, 2026)

added 37min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Illinois General Assembly
State legislature of Illinois, United States.

I am this person! Add a sourced quote Report comment

votes For and says:

reviewing assessments and adequacy of mitigations as part of the decision to deploy a frontier model or use it extensively internally; [...] assessing and managing catastrophic risk resulting from the internal use of its frontier models, including ri...
more AI Verified source (Jul 6, 2026)

added 39min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
OpenAI
AI research organization

I am this person! Add a sourced quote Report comment

votes For and says:

We also applied the method to challenging agentic rollouts, showing that it can extend beyond standard chat to more complex agent settings involving tool use, and can also be used for risk assessment before internal model deployments.
AI Verified source (Jun 16, 2026)

added 39min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Alex Mallen
AI safety researcher at Redwood Research.

I am this person! Add a sourced quote Report comment

votes For and says:

Risk reports commonly use pre-deployment alignment assessments to measure misalignment risk from an internally deployed AI. However, an AI [...] can develop widespread dangerous motivations during deployment. I think [...] AI companies and evaluators...
more AI Verified source (May 15, 2026)

added 37min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Anthropic
AI safety research company

I am this person! Add a sourced quote Report comment

votes For and says:

Before approving a new model for easy or broad internal use (e.g., making the model available via our most-used internal tools), we conduct a pre-deployment review process.
AI Verified source (2026)

added 39min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Frontier Model Forum
Frontier AI industry forum

I am this person! Add a sourced quote Report comment

votes For and says:

Frontier AI frameworks typically employ a two-stage risk assessment process carried out by internal assessment teams: Frontier Capability Assessments and Mitigations Assessments. [...] Developers then evaluate their effectiveness and whether the resi...
more AI Verified source (Aug 4, 2025)

added 37min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Zach Stein-Perlman
AI safety and governance researcher and writer.

I am this person! Add a sourced quote Report comment

votes For and says:

Do great risk assessment, especially via model evals for dangerous capabilities. Present risk assessment results and deployment plans to external auditors — perhaps make risk cases and safety cases for each planned deployment, based on threat modelin...
more AI Verified source (Oct 21, 2024)

added 37min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Government of the United Kingdom
United Kingdom central government

I am this person! Add a sourced quote Report comment

votes For and says:

Conducting thorough risk assessments before developing or deploying new models, complemented with continual monitoring (incorporating, for example, evaluation results, evidence from earlier models, and expert forecasts). [...] For example, at a given...
more AI Verified source (2023)

added 39min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Institute for AI Policy and Strategy
Think tank developing policy recommendations for frontier artificial intelligence, safety, and national security.

I am this person! Add a sourced quote Report comment

votes For and says:

Congress should establish a mandatory requirement to provide government access to, and evaluation of, frontier models before consequential internal deployment, to ensure visibility where the biggest risks are likely to emerge.
Unverified source (Sep 11, 2026)

added 54s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Institute for Progress
Nonpartisan think tank focused on innovation policy, economic growth, and technological progress.

I am this person! Add a sourced quote Report comment

votes For and says:

The developer’s publicly available frontier AI framework must describe how the developer reviews risk assessments in deciding to internally deploy the model; [...] the developer must periodically submit confidential reports to the Department of Comme...
more Unverified source (Aug 6, 2026)

added 52s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
METR
Nonprofit AI safety research organization that evaluates frontier AI systems.

I am this person! Add a sourced quote Report comment

votes For and says:

Most importantly, third-party evaluations triggered by public deployments are not designed to cover risks from internal use within the developer itself. [...] We believe that periodic third-party assessment of risks from developers’ internal use of A...
more Unverified source (May 19, 2026)

added 1min ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
UK AI Security Institute
United Kingdom government institute for evaluating advanced AI systems.

I am this person! Add a sourced quote Report comment

votes For and says:

As part of a larger body of work on understanding loss-of-control risks from frontier AI, we’ve been developing methodology to simulate how models internally deployed within AI companies might behave when performing and assisting with AI safety resea...
more Unverified source (Apr 27, 2026)

added 56s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Meta
Social media and AI company

I am this person! Add a sourced quote Report comment

votes For and says:

For Muse Spark, we conducted extensive safety evaluations before deployment. Because of its advanced reasoning capabilities, we evaluated the model before and after applying protections — testing not just for the most serious risks like cybersecurity...
more Unverified source (Apr 8, 2026)

added 44s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
New York State Senate
State legislative chamber, New York

I am this person! Add a sourced quote Report comment

votes For and says:

Before, or concurrently with, deploying a new frontier model, a frontier developer shall include in its public frontier AI framework summaries of assessments of catastrophic risks from the frontier model, the results of evaluations, and the extent to...
more Unverified source (Mar 27, 2026)

added 37s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Oxford Martin AI Governance Initiative
Oxford Martin research initiative studying governance and risk management for frontier artificial intelligence.

I am this person! Add a sourced quote Report comment

votes For and says:

Risk management standards emphasise that such decisions should follow a continuous and iterative process of risk evaluation and analysis, including reassessment after mitigations are implemented. [...] Because the most cutting-edge AI systems are oft...
more Unverified source (Feb 24, 2026)

added 46s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Center for Security and Emerging Technology
Georgetown University research center focused on emerging technology and national security.

I am this person! Add a sourced quote Report comment

votes For and says:

Risk management within AI companies is crucial for internal deployments of frontier AI systems for AI R&D specifically. [...] Further work is needed to develop best practices for risk management that explicitly cover the internal deployment setting, ...
more Unverified source (2026)

added 50s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
xAI
Artificial intelligence company developing the Grok family of frontier AI models.

I am this person! Add a sourced quote Report comment

votes For and says:

Pre-deployment reviews include assessing benchmark results (e.g., WMDP scores) and mitigation effectiveness. For internal use, we review catastrophic risks like oversight evasion before extensive rollout.
Unverified source (Dec 31, 2025)

added 42s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
California State Legislature
State legislative body

I am this person! Add a sourced quote Report comment

votes For and says:

A large frontier developer shall write, implement, comply with, and clearly and conspicuously publish on its internet website a frontier AI framework that ... [reviews] assessments and adequacy of mitigations as part of the decision to deploy a front...
more Unverified source (Sep 29, 2025)

added 38s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Apollo Research
AI safety research organization

I am this person! Add a sourced quote Report comment

votes For and says:

Expand Frontier Safety Policies (FSPs) to explicitly cover internal AI deployment, with specific evaluations for scheming detection and control before internal use and appropriate mitigations.
Unverified source (Apr 17, 2025)

added 40s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Buck Shlegeris
CEO of Redwood Research; AI safety researcher focused on AI control

I am this person! Add a sourced quote Report comment

votes For and says:

I think labs should probably make safety cases for models capable of causing catastrophe in two parts, matching those two classifications of catastrophe: To prevent catastrophe without rogue deployment, we ensure: While our safety measures are in pla...
more Unverified source (Jun 3, 2024)

added 47s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
DeepSeek
Chinese artificial intelligence company developing open-weight frontier models

I am this person! Add a sourced quote Report comment

votes For and says:

DeepSeek takes the potential risks of AI technology applications very seriously. We strictly comply with legal and regulatory requirements and take reasonable measures to continuously enhance model safety throughout the entire lifecycle of model deve...
more Unverified source

added 41s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Americans for Responsible Innovation
AI policy advocacy nonprofit

I am this person! Add a sourced quote Report comment

votes For and says:

To promote responsible frontier innovation, federal frontier governance should cover AI used internally by a developer in addition to any public or external offerings; there is already evidence that dangerous capability thresholds can be reached by i...
more Unverified source

added 51s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Google DeepMind
AI research company.

I am this person! Add a sourced quote Report comment

votes For and says:

The categories we are considering span from those containing low-risk models, in which our current processes already operate, to those containing models that may pose greater risk, where additional mitigations may be recommended, such as subjecting m...
more Unverified source

added 55s ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate