Demis Hassabis Proposes an Independent Standards Body for Frontier AI

Introduction

Google DeepMind co-founder and CEO Demis Hassabis has proposed a new independent body for testing and overseeing the world’s most capable AI models.

The proposed organization would sit between frontier AI laboratories and the U.S. government. It would develop technical evaluation standards, test advanced models before release, update benchmarks as capabilities improve, encourage stronger security practices, and coordinate responses when serious vulnerabilities appear after deployment.

Hassabis outlined the plan in an article titled “A Framework for Frontier AI and the Dawning of a New Age.” TechCrunch reported on the proposal on July 14, 2026, emphasizing its unusual structure: an industry-funded, technically staffed organization modeled partly on the Financial Industry Regulatory Authority, or FINRA.

The plan begins with voluntary cooperation. Frontier AI developers would provide qualifying models for assessment as much as 30 days before release. If the evaluation system proved reliable, passing the review could later become a requirement for deploying a frontier model in the U.S. market.

Why Hassabis Believes a New System Is Needed

Hassabis argues that highly capable AI systems may arrive faster than governments, researchers, and society can develop dependable ways to evaluate them.

The concern is not limited to whether a model produces offensive text or makes ordinary factual mistakes. Frontier systems may develop capabilities that affect:

• Cybersecurity
• Biological research
• Chemical or radiological risk
• Autonomous tool use
• AI research automation
• Persuasion and manipulation
• Deceptive behavior
• Model-weight security
• Critical infrastructure
• National security

A model can improve significantly between regulatory cycles. Evaluations that were difficult six months ago may become saturated and stop revealing meaningful differences between systems.

Traditional rulemaking is often slow. Frontier-model development is not.

Hassabis therefore proposes an organization whose assessment methods can change at approximately the same pace as the technology.

The Core Proposal

The proposed standards body would be a federally overseen public-private organization or self-regulatory body.

It would be technically independent in its day-to-day work while cooperating with government agencies and national laboratories on issues connected to public safety and national security.

Its main responsibilities would include:

1. Defining which models qualify as frontier-class.
2. Developing and maintaining evaluation protocols.
3. Reviewing qualifying models before release.
4. Testing dangerous capabilities and safeguard reliability.
5. Establishing security and transparency practices for frontier laboratories.
6. Supporting an ecosystem of external evaluators and auditors.
7. Responding to serious post-release vulnerabilities.
8. Coordinating stronger interventions if risks become unacceptable.

The standards body would not regulate every small model, startup, university project, or ordinary AI application.

Its scope would be limited to systems that pass capability thresholds indicating that they may create unusually serious risks.

How a Model Would Become “Frontier-Class”

Hassabis’s framework does not define a frontier model solely by parameter count, training cost, company size, or brand recognition.

Instead, the proposed standards body would maintain a set of capability benchmarks.

A model that crosses specified thresholds would be classified as frontier-class. An organization developing such a model would then be treated as a frontier laboratory.

This capability-based approach matters because model architecture and efficiency are changing quickly.

A smaller or cheaper system may eventually match capabilities that once required a much larger training run. A fixed threshold based only on compute could become outdated.

The proposed benchmarks would therefore be:

• Capability-based
• Regularly revised
• Designed around high-risk domains
• Retired when they become too easy
• Replaced when models begin to overfit or saturate them
• Applied regardless of whether a model is open or closed

The classification could also apply to models developed outside the United States if they were deployed in the U.S. market.

A Phased Pre-Release Review Process

The proposal uses a phased implementation rather than moving immediately to mandatory licensing.

Phase 1: Voluntary Model Sharing

Frontier laboratories would voluntarily provide qualifying models to the standards body before release.

The review window could last up to 30 days.

During that period, evaluators would assess the model’s capabilities, weaknesses, safeguards, and potential national-security risks.

The initial voluntary stage would help the organization:

• Build technical expertise
• Test evaluation methods
• Improve confidentiality procedures
• Learn how much time different assessments require
• Establish secure model-access systems
• Identify benchmark gaps
• Build trust with developers and external evaluators

Phase 2: Formal Market Requirement

Once the assessment process was shown to be technically effective and operationally reliable, the framework could become formal.

Frontier models would then need to pass the process before they could be deployed in the United States.

This would move pre-release testing from a voluntary industry practice toward a defined market-access requirement.

Phase 3: Post-Release Monitoring

Approval would not end the standards body’s involvement.

Frontier laboratories would be expected to work with the organization on critical vulnerabilities discovered after launch.

That could include:

• New jailbreak methods
• Model-weight leakage
• Unexpected agent behavior
• Dangerous tool-use patterns
• Benchmark failures
• Safeguard regressions
• Newly discovered cyber or biological capabilities
• Risks that appear only under large-scale deployment

The framework therefore treats safety as a continuing lifecycle rather than a single approval event.

What the Standards Body Would Test

Hassabis calls for rigorous scientific testing across domains where advanced capabilities could create severe harm.

The initial evaluation areas would likely include:

Evaluation Area	Example Questions
Cybersecurity	Can the model discover, exploit, or automate serious software vulnerabilities?
Biological risk	Can it meaningfully assist advanced work that could increase biological danger?
Chemical and radiological risk	Does it provide capabilities beyond widely available information?
Agentic behavior	Can it plan, use tools, persist, or bypass restrictions over long tasks?
Deception	Does it conceal goals, manipulate evaluators, or behave differently when monitored?
Safeguard robustness	Can adversarial users bypass policies through jailbreaks or multi-turn attacks?
Model security	Could valuable model weights or sensitive system information be stolen?
AI research capability	Can the model accelerate the development of more capable AI systems?

The purpose would not be to claim that every risk can be measured perfectly.

The objective is to create evidence strong enough to support deployment decisions and identify where additional safeguards are needed.

Evaluations Would Need Constant Updates

One of the most important parts of the proposal is the requirement to revise tests frequently.

Hassabis suggests that evaluation suites might initially be updated every quarter.

This is necessary because frontier benchmarks can become unreliable in several ways:

• Models reach near-perfect scores.
• Training data contains benchmark answers.
• Developers optimize directly for the test.
• The benchmark no longer reflects real-world use.
• New model capabilities create risks the test was not designed to measure.
• Agentic systems learn to exploit weaknesses in the evaluation environment.

The standards body would begin by consulting frontier laboratories, which possess significant technical knowledge about current systems.

Over time, however, it would need to develop its own private, held-out evaluations.

These tests would not be fully visible to developers before assessment. That would reduce the risk of models being specifically trained to pass known benchmarks without demonstrating broader safety.

Why Independent Held-Out Tests Matter

Public benchmarks are useful for research and transparency, but they are not sufficient for high-stakes certification.

When developers know every evaluation question, they may:

• Train directly on the test set
• Adjust prompts for the benchmark
• Add narrow filters around known tasks
• Optimize model routing for the scoring method
• Produce strong benchmark results without improving general behavior

An independent evaluator can reduce these problems by maintaining confidential tasks and regularly introducing new ones.

The process resembles security testing more than a conventional leaderboard.

A credible evaluation program would need:

• Secure model access
• Version control
• Reproducible environments
• Clear scoring rules
• Independent test design
• Audit logs
• Adversarial testing
• Multiple expert reviewers
• Procedures for handling uncertain results
• A process for appeals and retesting

Without these safeguards, the standards body could become another benchmark publisher rather than a meaningful oversight institution.

Why Hassabis Uses FINRA as the Model

FINRA is a private, not-for-profit self-regulatory organization for U.S. broker-dealers.

It is funded by industry fees but operates under oversight from the Securities and Exchange Commission.

FINRA writes and enforces rules for its members, conducts examinations, monitors market activity, administers qualification processes, and responds to emerging risks.

Hassabis does not suggest copying every part of financial regulation into AI.

The relevant structural features are:

• Industry funding
• Government supervision
• Specialist staff
• Rule development
• Technical monitoring
• Membership or qualification requirements
• The ability to adapt to changing risks
• A mixture of public and industry participation

A frontier AI organization would require expensive computing infrastructure and highly specialized staff.

Industry funding could provide the resources needed to recruit model researchers, cybersecurity specialists, biological-risk experts, evaluation engineers, and infrastructure teams.

Federal oversight would be intended to prevent the organization from becoming a private club controlled entirely by the companies it evaluates.

Proposed Governance

Hassabis suggests that the board should include independent technical experts and representatives from the open-source AI community.

This is important because frontier AI governance affects organizations with very different development models.

Closed-model laboratories may provide access through secure interfaces without releasing weights. Open-model developers may publish model parameters, code, and technical details.

A standards body that only represented the largest proprietary labs could create rules that unintentionally or deliberately disadvantage smaller organizations.

A more balanced structure could include:

• Independent AI researchers
• Cybersecurity experts
• Biological and chemical safety experts
• Open-source representatives
• Consumer and civil-society voices
• National laboratory personnel
• Government observers
• Frontier-lab technical staff
• Evaluation researchers
• Infrastructure and security specialists

The exact voting rights and conflict-of-interest rules would be critical.

Industry expertise is necessary to understand the technology, but excessive industry control could weaken public trust.

Best Practices for Frontier Laboratories

The proposal goes beyond model testing.

Organizations classified as frontier laboratories would be encouraged to adopt stronger operational practices.

These could include:

• Publishing detailed model cards
• Maintaining strong internal cybersecurity
• Protecting model weights
• Vetting personnel in sensitive roles
• Funding safety and security research
• Documenting evaluation results
• Reporting significant incidents
• Applying digital watermarks to generated media
• Maintaining secure deployment controls
• Cooperating on post-release vulnerabilities

These requirements reflect an important reality: frontier AI risk does not begin and end with model behavior.

A safe model can still create risk if its weights are stolen, its system prompt is exposed, its tools are misconfigured, or its deployment environment lacks proper controls.

Third-Party Evaluators Would Still Be Needed

The standards body would not need to perform every test internally.

Hassabis proposes supporting an ecosystem of independent evaluators and auditors.

Different organizations could specialize in:

• Cybersecurity evaluations
• Biological-risk assessments
• Agent autonomy
• Manipulation and persuasion
• Interpretability
• Model-weight protection
• Red teaming
• Open-model testing
• Watermarking and provenance
• Post-deployment monitoring

The central body could define requirements and coordinate the process while approved specialists conduct particular evaluations.

This approach may be more scalable than attempting to build every form of expertise inside one organization.

It also reduces dependence on a single evaluator whose methods may contain unknown blind spots.

How the Proposal Relates to Existing U.S. Institutions

The United States already has government organizations working on AI testing and standards.

The National Institute of Standards and Technology operates the Center for AI Standards and Innovation, or CAISI.

CAISI works with industry on voluntary standards, develops evaluation methods, and conducts assessments related to cybersecurity, biosecurity, chemical weapons, national security, and foreign AI systems.

Hassabis’s proposal overlaps with this work but would create a distinct structure.

A possible division of responsibilities could look like this:

Institution	Possible Role
CAISI and NIST	Measurement science, public standards, government evaluations, national-security coordination
Proposed standards body	Pre-release assessment process, frontier-lab requirements, technical certification, ongoing industry oversight
Federal agencies	Legal authority, national-security decisions, enforcement, export controls
National laboratories	Specialized testing, secure infrastructure, scientific expertise
Third-party evaluators	Domain-specific audits, red teaming, benchmark development
Frontier laboratories	Secure model access, technical documentation, remediation, post-release monitoring

This arrangement would require careful coordination to avoid duplicated tests, conflicting requirements, and unclear authority.

Relationship to Google DeepMind’s Frontier Safety Framework

Google DeepMind already uses an internal Frontier Safety Framework.

The framework identifies critical capability thresholds and connects them to evaluation, security, and deployment measures.

Its risk areas have included:

• Cybersecurity
• CBRN capabilities
• Harmful manipulation
• Machine-learning research and development
• Misalignment

DeepMind’s framework also uses safety cases and deployment mitigations for models that reach defined capability levels.

Hassabis’s new proposal would extend this general logic beyond one company.

Instead of each frontier laboratory defining and reviewing its own thresholds, an external organization would create shared tests and minimum requirements.

Potential Strengths of the Proposal

Technical Expertise

A specialized organization could recruit people who understand advanced models deeply enough to evaluate them.

General regulatory agencies often struggle to match the technical depth of frontier laboratories.

Faster Adaptation

A self-regulatory structure may be able to update benchmarks and procedures faster than legislation.

Shared Standards

Common evaluations could make model-safety claims easier to compare across developers.

Pre-Release Access

Testing before deployment gives evaluators more time to identify severe weaknesses before they reach millions of users.

Independent Testing

Held-out tests and third-party auditors could provide evidence beyond company-published model cards.

Limited Scope

Exempting ordinary, non-frontier models could reduce unnecessary burdens on startups, universities, and smaller open-source projects.

International Foundation

A U.S.-led system could provide a starting point for compatible international evaluation standards.

Important Questions the Proposal Does Not Fully Resolve

The plan is detailed, but several design questions remain open.

Who Decides What Counts as Frontier AI?

Capability thresholds can determine which companies face expensive testing and possible release delays.

The process for setting those thresholds must be transparent, evidence-based, and open to challenge.

How Independent Can an Industry-Funded Body Be?

Industry funding can provide expertise and compute, but it can also create conflicts of interest.

Strong government oversight, public-interest governance, financial transparency, and external review would be needed.

Can a 30-Day Review Be Completed Reliably?

Some evaluations require specialized environments, expert review, long-running agent tests, or repeated experiments.

The standards body would need enough staff and compute to avoid becoming a release bottleneck.

How Would Open Models Be Tested?

A model whose weights will be released creates different risks from a hosted API.

Once weights are public, many deployment safeguards can be removed.

Evaluation requirements may need to distinguish between open weights, restricted weights, and closed hosted systems.

What Happens When Results Are Uncertain?

Frontier evaluations often produce ambiguous evidence.

The framework would need rules for conditional approval, additional testing, deployment restrictions, appeals, and independent review.

How Would Enforcement Work?

A self-regulatory body cannot automatically impose legal penalties on companies outside its jurisdiction.

Formal authority would require legislation, agency action, contractual requirements, procurement rules, or market-access conditions.

Could the System Protect Incumbents?

Large AI companies can afford extensive evaluations and compliance teams.

Smaller developers may struggle, even if their models meet the same capability thresholds.

The framework would need to prevent safety requirements from becoming an anticompetitive barrier.

When the Framework Could Become More Restrictive

Hassabis argues that the system should be adjustable.

If evaluations show that frontier models are creating more serious risks, oversight could be strengthened.

Possible escalations might include:

• Longer review periods
• Additional security requirements
• Restricted deployment environments
• Mandatory human supervision
• Limits on tool access
• Controlled access to model weights
• Stronger incident-reporting duties
• Temporary release restrictions
• Coordinated pauses for specific capability categories

The most serious option would be coordinating a slowdown among frontier laboratories if the evidence justified it.

Such an intervention would be politically and technically difficult. It would require clear thresholds, international coordination, enforcement mechanisms, and credible evidence that continuing at the same pace posed an unacceptable risk.

The International Problem

Frontier AI is developed and deployed across national borders.

A U.S. standards body could influence access to the American market, but it could not independently govern every laboratory worldwide.

Hassabis’s proposal therefore treats U.S. leadership as a starting point rather than the final system.

International compatibility would be needed around:

• Capability definitions
• Evaluation methods
• Confidential model access
• Incident reporting
• National-security risks
• Open-weight releases
• Model provenance
• Export and deployment controls
• Mutual recognition of test results
• Cross-border auditing

Without coordination, companies may face inconsistent standards or choose the jurisdiction with the weakest rules.

At the same time, a single global authority may be politically unrealistic.

A network of national and regional evaluation bodies using shared technical standards may be more achievable.

What the Proposal Means for Frontier AI Developers

A functioning standards body would change the model-release process.

Today, leading laboratories mainly publish their own model cards, system cards, safety frameworks, and selected evaluation results.

Under the proposed system, frontier developers could need to prepare for:

1. Confidential pre-release model access
2. Independent capability testing
3. Held-out benchmarks
4. External security review
5. Standardized incident reporting
6. Evidence of internal cybersecurity
7. Post-release remediation obligations
8. More detailed technical documentation
9. Possible deployment conditions
10. Formal approval before U.S. market access

Safety evaluation would become a defined release gate rather than an internal practice chosen by each company.

常见问题

What did Demis Hassabis propose?

He proposed an independent, U.S.-led standards body for evaluating frontier AI models. It would develop technical tests, review advanced models before release, promote security practices, and coordinate responses to serious vulnerabilities.

Would the organization be part of the U.S. government?

The proposed structure would resemble a federally overseen public-private partnership or self-regulatory organization. It could be funded largely by the AI industry while operating under government supervision.

Why is FINRA being used as a model?

FINRA is an industry-funded self-regulatory organization that oversees U.S. broker-dealers under SEC supervision. Hassabis believes a similar structure could combine technical expertise, adaptable rules, industry funding, and public oversight.

How long would frontier models be reviewed before release?

The initial proposal allows voluntary sharing with the standards body for up to 30 days before release. The exact review time could vary depending on the model and evaluation requirements.

Would every AI model need approval?

No. The framework would apply to models that cross regularly updated frontier-capability thresholds. Smaller models from startups, universities, and other organizations would generally remain outside the process unless their capabilities reached the defined level.

Would open-source AI models be covered?

Hassabis says the framework should apply to frontier-class systems whether they are open or closed and regardless of country of origin. Open-weight models may require different safeguards because deployment controls can be removed after release.

What risks would the body evaluate?

The proposal highlights cybersecurity, biological threats, agentic behavior, deception, safeguard bypass, and other high-risk capabilities. The test suite would need regular updates as new risks and capabilities appear.

Does the United States already evaluate advanced AI models?

Yes. NIST’s Center for AI Standards and Innovation works on testing, standards, and assessments involving commercial and frontier AI. The proposed body would add a more formal, independent pre-release review and industry-oversight structure.

相关工具

• Inspect AI: An open-source framework from the UK AI Security Institute for structured and repeatable frontier-model evaluations.
• NIST AI Risk Management Framework: A voluntary framework for governing, mapping, measuring, and managing AI risks.
• NIST Center for AI Standards and Innovation: The U.S. government center for AI testing, measurement science, standards, and national-security evaluations.
• Google DeepMind Frontier Safety Framework: DeepMind’s internal framework for identifying and mitigating severe risks from frontier models.
• Inspect Evals: A collection of reusable evaluations built on the Inspect AI framework.

Related Links

• A Framework for Frontier AI and the Dawning of a New Age: Demis Hassabis’s complete proposal.
• About FINRA: FINRA’s official explanation of its structure, funding, responsibilities, and SEC oversight.
• Center for AI Standards and Innovation: Official information on CAISI’s evaluations, standards work, and national-security role.
• Google DeepMind Frontier Safety Framework Update: DeepMind’s approach to capability thresholds, safety cases, security, and deployment mitigations.
• UK AI Security Institute Evaluation Approach: The UK government’s official approach to pre- and post-deployment advanced-AI testing.
• NIST Generative AI Risk Management Profile: Guidance for managing risks across the generative-AI lifecycle.
• International Network Joint Statement on Advanced AI Risk Assessment: Shared principles for evaluating advanced AI systems internationally.

Summary

Demis Hassabis is proposing a technically specialized, independently operated standards body that would review the most capable AI models before release. The organization would begin with voluntary cooperation, build reliable evaluation protocols, and could later become a required gate for deploying frontier models in the United States.

The proposal attempts to combine industry expertise and funding with government supervision, independent testing, held-out benchmarks, third-party auditors, and continuing post-release monitoring.

Many implementation questions remain, including governance, enforcement, open-model treatment, evaluation capacity, conflicts of interest, and international coordination. Even so, the proposal is more concrete than a general call for AI regulation because it identifies an institutional structure, review process, scope, and technical responsibilities.

The central idea is straightforward: frontier AI safety should be tested by an organization with enough independence, expertise, compute, and authority to challenge the companies building the models.