SUMMARY
- EU law allows the Commission to compel employee-level access to AI models for specified evaluations, backed by fines and powers to restrict models’ availability.
- The EU’s voluntary code sets expectations for privileged outside scrutiny, while Brussels pays evaluators to develop tests and assessment methods.
- The campaign for permanent embedding extends into the EU’s advisory network. Apollo Research, whose leaders sit on its scientific panel, wants governments to make embedded evaluations compulsory.
The European Union can compel AI companies to give outside assessors access at every level granted to their own employees. Its code expects privileged external scrutiny, and Brussels pays evaluators to develop the tests regulators will use. The push to make that oversight permanent now includes an organization whose leaders advise the EU on AI enforcement.
EU Law Allows Employee-Level Access
Under the AI Act’s Article 92, the Commission can evaluate general-purpose models to investigate systemic risks or assess compliance when the information supplied is insufficient. It can appoint independent experts, including scientific panel members, and require companies to provide access to their models.
A July 2026 implementing regulation allows that access to include source code, model weights, internal systems and hosting infrastructure, depending on the evaluation’s objectives. Evaluators can inspect and modify system state. The regulation also allows access at every level granted to the company’s own employees.
The Commission can require companies to disable logging that would track its access. It can appoint scientific panel members to evaluations without a further selection procedure. The rules also require independence checks and confidentiality commitments, and allow companies to raise reasoned objections to experts.
These powers apply whether or not a company signs the voluntary code. Outside experts can already gain employee-level access under compulsory EU powers for specified evaluations. Permanent embedding would extend that access into an ongoing role inside the laboratory.
Under Article 93, the Commission can require companies to mitigate risks and, where appropriate, restrict a model’s availability or require its withdrawal or recall. Article 101 allows fines for infringements including failure to provide required access, up to the greater of €15 million or 3 per cent of worldwide annual turnover.
What the Voluntary Code Expects
The General-Purpose AI Code of Practice sets out the Commission’s preferred route for demonstrating compliance with binding AI Act obligations. Its signing guidance allows other adequate means of compliance, but says non-signatories may face additional requests for information and access. The code is voluntary, but companies have a clear regulatory incentive to follow it.
The code’s safety and security provisions call for qualified independent external evaluations of systemic-risk models. Exceptions include models shown to be comparably safe and cases where reasonable efforts to secure an evaluator have failed. Evaluators’ access can include training data, system prompts, activations, gradients and model reasoning.
After a model reaches the market, the code also envisages free access for independent evaluators to capable versions with minimal guardrails. Privileged outside scrutiny is therefore part of the oversight Brussels expects throughout a model’s development and use.
Helping Write the Tests
The Commission’s technical-assistance tender covered six areas, including cyber risks, loss of control, harmful manipulation and agentic systems. The contractors help develop tests and thresholds that regulators can use to assess companies’ compliance.
The EU also pays external organizations to develop evaluation methods. METR’s own disclosures identify a European AI Office contract for work on loss-of-control risks, which it says provides a small share of its funding. FAR AI announced that it had been selected to lead an EU AI Act consortium on chemical, biological, radiological and nuclear risks, and to participate in work on harmful manipulation.
The EU scientific panel advises the AI Office and national authorities, identifies emerging risks, contributes to evaluation methods and supports enforcement. Its members help shape both the risks regulators investigate and the methods they use.
The EU’s published panel membership includes Marius Hobbhahn and Charlotte Stix. Apollo Research’s leadership page lists Hobbhahn as its chief executive and co-founder, and Stix as its chief governance officer. Hjalmar Wijk, chief scientist at Model Evaluation and Threat Research, or METR, also sits on the panel.
Apollo and METR were involved before the current embedding debate. At an EU AI Office workshop in April 2025, Hobbhahn and METR chief executive Beth Barnes presented on loss-of-control and persuasion risks. Stix and METR’s Max Hasin joined the concluding discussion on evaluation practices. More than 80 participants from over 50 organizations attended.
FFO’s investigation of the emerging AI safety blob traced the developers, donors, evaluators and policy advocates seeking greater control over frontier AI. Some of those evaluators now work with European regulators as well as the companies they assess.
FFO previously examined the RAND–METR AI safety collaboration and the funding and institutional ties surrounding these evaluators. METR’s EU contract gives it another role in oversight, helping a regulator develop the methods used to judge AI systems.
RAND Corp, Backed by $330M in Federal Funds, Teams up with METR on AI Safety
FFO’s reporting on the DSA enforcement network documented how EU regulation supports a market for contractors and outside expertise. AI evaluation follows a similar pattern, with Brussels imposing obligations and hiring outside organizations to help enforce them.
These evaluations also cover how AI handles political information. The code treats harmful manipulation through distorted beliefs, persuasion and deception as a risk, including its effects on democratic processes and fundamental rights.
Organizations helping define manipulation tests can influence which behavior regulators consider harmful and which restrictions companies are expected to introduce. Their judgments may therefore affect what AI systems say and how users can interact with them.
The Push for Permanent Embedding
Anthropic chief executive Dario Amodei proposes permanent embedded evaluators in “We Must Pace the Frontier”, complete with desks, badges, laptops and internal permissions. He calls for frontier laboratories to work with governments to formalize the arrangement and names METR as an example of an evaluator.
Anthropic has announced Accenture as its first embedded evaluation partner, with Accenture’s Faculty business leading the work. Anthropic will pay for the engagement. It is also discussing pilots with METR and other nonprofit evaluators using their own funding. Accenture was already a commercial partner, having announced a dedicated Anthropic business group with the company in December 2025.
An embedded evaluator may depend on the company it assesses for payment, infrastructure and continued access. Where the two also have a commercial partnership, the evaluator’s findings could affect that business relationship.
The AI Evaluator Forum’s September embedded-evaluation letter calls for protected funding, editorial autonomy and access to privileged information, staff and physical spaces. It also says evaluators should avoid other significant commercial relationships with the laboratories they scrutinize.
In a separate September essay, Apollo argues that occasional external tests cannot reliably uncover the dangers it is concerned about. It wants continuing access to staff, systems and company processes during training and internal deployment.
In its September statement, Apollo argues that voluntary scrutiny will not be enough and that governments should eventually require embedded evaluators by law. As mentioned, its chief executive and chief governance officer sit on the EU’s scientific panel, putting leaders of an organization campaigning for compulsory embedding inside Brussels’ advisory network.
The EU already has the power to compel deep internal access and a network of outside organizations helping develop its oversight methods. Permanent embedding would extend those arrangements into continuing supervision inside AI laboratories. Advocates for that expansion are represented in the institutions advising Brussels on how to oversee the industry.



