Dario Amodei, co-founder and CEO of Anthropic, speaks to “The Circuit with Emily Chang” at Anthropic’s headquarters in San Francisco, California, on Thursday, April 30, 2026.
Jason Henry | Bloomberg | Getty Images
More than 100 artificial intelligence experts and evaluators have united to warn that they are not getting the resources and protections they need to test the safety of AI technology. Artificial intelligence technology is facing increased scrutiny due to insider concerns about the potential dangers of frontier models.
“We’re just trying to really demonstrate common ground on fundamental principles and ensure that independent oversight can be a meaningful tool for broadly managing AI risks,” Konrad Stoss, president of the AI Evaluators Forum Consortium, which sponsored the letter, said in an interview.
The group released an open letter on the issue on Friday, shared exclusively with CNBC. Signatories include AI luminaries like Jeffrey Hinton and members of organizations such as Johns Hopkins University, Stanford University, and the nonprofit evaluation organization METR. The letter says it wants to force underlying model providers to ensure that third-party AI evaluators are allowed the “scientific objectivity, transparency, independence, and robust protections” they need to perform their work effectively and reliably.
Stoss said this is part of an effort to hold foundation model companies accountable for recent promises to support more thorough third-party AI safety testing.
The niche evaluator community has been thrust into the spotlight since Anthropic CEO Dario Amodei announced the idea over the weekend to give some evaluators “employee-like access” to inspect and audit cutting-edge underlying models and their development processes. Some industry leaders have called for the government to regulate AI development to prevent it from spiraling out of control, but President Donald Trump and his former AI czar David Sachs have staunchly opposed such efforts.
Stoss said the coalition was not “advocating any particular method” to ensure that AI models are developed safely, but wanted to ensure that “ground rules” and “higher standardization” were at least established for evaluators and others working independently from major research institutions.
Stoss said Amodei’s proposal appears to include providing evaluators with significantly more access than they have previously enjoyed. Stoss said such a scenario could include the creator of the underlying model giving third-party evaluators access to company computers, allowing them to speak openly with employees and “view sensitive internal data and non-public systems.”
“Having that type of access would give us greater confidence and certainty about the actual risk, especially for systems that we use internally and have not released,” Storrs said. He cited the unreleased OpenAI model used in the Hugging Face attack.
OpenAI CEO Sam Altman space x Elon Musk and microsoft CEO Satya Nadella has publicly supported Amodei’s proposal, but has yet to address the logistical issues that come with such an undertaking, such as which AI evaluators would be chosen and how deeply they would examine the closely guarded technology.
The signatories want the evaluators’ work to be independent of companies, to increase transparency about technology and to be “protected from retaliation from the companies they work with,” the letter said.
integrated power
Vinh Nguyen, a senior AI fellow at the Council on Foreign Relations and former chief AI officer at the National Security Agency, said independent evaluators are needed to unearth critical information that can help mitigate potential security failures and economic disasters.
“When a small number of powerful laboratories control capabilities that can compromise cybersecurity, critical infrastructure, and the systems underlying our national security and economy, governments and the public cannot rely on the laboratories’ own narratives about what is safe,” Nguyen, who signed the letter, said in a statement.
Stoz said third-party evaluators are not intended to be “a replacement for internal evaluation efforts, let alone mitigating issues discovered by developers.” He acknowledged that the Foundation model companies may ignore open letters and calls to action, but said their credibility was at stake.
“There are very few groups that actually have enough technical credibility and the scale and capacity to do this kind of work,” he said.
Read the full letter below and click here for a link to the list of signatories.
Minimum conditions for embedding an evaluator
We, the undersigned, encourage to see cutting-edge AI companies seek to incorporate third-party organizations to assess rapidly escalating AI capabilities and risks. We believe that all frontier AI companies should incorporate assessors to independently assess AI risks, including assessing critical incidents that harm the system itself or the real world, and the company’s training, deployment, monitoring, operations, and safety practices.
Incorporating reliable third-party evaluations requires scientific objectivity, transparency, independence, and robust protections against interference from the companies being evaluated, including at least:
Frontier AI companies should rely on evaluators who have meaningful independence, maintain full editorial control, and disclose and mitigate potential conflicts of interest. This includes, at a minimum, that the embedded evaluation organization should not be owned or controlled by a frontier AI company, should not have any other significant commercial transactions, and should not accept any form of payment or other compensation in response to the evaluator’s findings.
Frontier AI companies will need to incorporate different perspectives and areas of expertise, including multiple rating organizations across different priority risk areas, each with deep relevant technical expertise, and allowing and encouraging raters to share how their conclusions vary between raters and between raters and employees.
Embedded evaluators must be transparent, including transparency about their methods and findings, the nature of their access, and the broader terms of their evaluation. Frontier AI companies should actively promote this transparency, including by limiting the scope of non-disclosure agreements. It should also allow evaluators to have prompt and unfiltered communication with the company’s board of directors and other privileged supervisory bodies, as well as the publication of findings and evidence, subject only to a time-limited redaction process limited to the protection of intellectual property, confidential customer information, and important interests in personal privacy, security, and public safety.
Embedded evaluators should be protected from retaliation from embedded companies for selecting reasonable valuation methods, discovering information, or drawing conclusions that are unpleasant to those companies. This includes funding mechanisms that provide reasonable protection against retaliation lawsuits and confidence that funding will continue in such cases.
Frontier AI companies should allow built-in evaluators the same access as their own highly privileged employees for evaluation purposes, with exceptions to protect sensitive data belonging to the company’s customers or other third parties. This includes access to the same relevant systems, data, tools and physical space available to senior internal employees responsible for conducting comparable risk assessments, as well as open and direct one-on-one communication with relevant staff.
This list is not comprehensive, and such conditions to ensure reliable assessments will increasingly need to be standardized, codified and enforced. One example is the set of terms defined in the AEF-1 standard, which has already been adopted early, but more work is needed to ensure that built-in evaluators are effective and meaningful.
Embedded assessments cannot address all oversight needs and should be treated as a complement to, rather than a replacement for, broader efforts by frontier AI companies to expand external oversight, including increased public transparency and additional broader forms of access for independent researchers.
WATCH: Public release of AI technology ‘couldn’t be worse’.
