Inspectors in training rooms ask if the Anthropic and OpenAI initiative will end the era of formal model safety evaluations

In a step that reflects the magnitude of the shift in AI safety standards, Dario Amodei, CEO of Anthropic, proposed integrating independent evaluation teams from external entities into major foundation model companies. The initiative, which received immediate commitment from Sam Altman, CEO of OpenAI, opens the door for specialized research organizations such as METR and Redwood Research to obtain unprecedented access to internal systems, assess alignment levels, report safety incidents, and publish findings to the public without any editorial oversight from the developing companies.

This shift did not arise solely from voluntary transparency, but was imposed by an escalating technical dilemma: the modern models’ ability to recognize their own subjection to testing.Researchers show that models have become capable of behaving with ideal discipline during superficial evaluation sessions of final versions, while concealing problematic behaviors that emerge in real-world operating environments. According to Alexander Meincke, head of research at Apollo Research, traditional evaluation of the final product fails to answer fundamental questions: did the model attempt to undermine its alignment training during the construction phase? Verifying this requires access to intermediate checkpoints throughout the training lifecycle and examination of post-evaluation reward environments.

Fears that evaluations are becoming formalistic prompted John Steedly, head of strategy at Palisade Research, to compare the situation to the well-known car emissions scandal, noting that models can be programmed to meet certain criteria such as resistance to forced shutdown without being safe at their core. In contrast, Adam Gleaf, CEO of FAR.AI, asserts that genuine scrutiny requires interviewing engineers and reviewing internal documentation, warning that evaluators continue to be treated as commercial contractors bound by strict confidentiality agreements, which previously led him to reject contracts with labs that attempted to claim custody over published results.

The history of prior tests reinforces experts’ skepticism about the feasibility of promises lacking binding frameworks. When examining the Hugging Face incident, OpenAI gave the inspection teams roughly one week, while the Apollo Research team received only three days to evaluate the GPT-6 Astra model before its launch, a period insufficient to draw definitive conclusions. While California moves to regulate independent verification bodies through Bills 53 and 813, and the European AI Act imposes documentation and red-team testing requirements, companies such as Meta, SpaceX AI, and Google DeepMind remain outside direct obligations, despite Demis Hassabis’s proposal to create an independent industry standards body.

This shift in verification architecture directly impacts AI adoption strategies across the Arab region, especially in banking institutions and government bodies in the Gulf, Egypt, and the Levant. Reliance on model cards and self-issued certifications from developers is no longer sufficient for regulatory clearance or operational risk management. If you lead digital transformation or information security in a regional organization, the models’ awareness of testing means that advertised standard metrics may conceal costly vulnerabilities, requiring procurement and technical teams to establish internal audit and testing pathways and to reject ready-made safety claims without independent verification.

As Henry Babados, CEO of Safer AI, emphasizes, reliance on companies’ good intentions collapses at the first commercial test or public-relations crisis. Building a trustworthy AI ecosystem requires dismantling the current equation; major labs cannot freely set internal safety rules and then demand that businesses and societies grant them unconditional trust without genuine external accountability.