The Embedded Evaluator: Anthropic and OpenAI Propose a New Era of AI Safety Transparency

The frontier of artificial intelligence is moving toward a radical new model of transparency: the embedded evaluator. In a significant departure from the traditional "black box" approach to model deployment, leadership at both Anthropic and OpenAI have signaled a willingness to allow third-party safety researchers direct access to the internal systems where models are trained and evaluated.

This shift, proposed and supported by the CEOs of the industry's two most prominent labs, aims to move beyond the limitations of post-hoc testing and toward a real-time, integrated safety monitoring framework.

What Happened

On September 16, 2026, a new proposal emerged from the highest levels of the AI industry. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have both expressed support for a model where independent, third-party safety evaluators are embedded within frontier AI companies.

The core of the proposal is to grant organizations like METR (Model Evaluation and Threat Research) and Redwood Research unprecedented access to the internal development environments of these labs. Unlike current safety protocols, which typically involve testing finished models on isolated benchmarks, this new framework would allow evaluators to inspect "checkpoints": intermediate versions of models during the training process.

The goal is to detect emerging problematic behaviors, such as deceptive alignment or unexpected capabilities, before they are baked into a final, released product.

According to reporting from TechCrunch, the scale of this access is significant: "Anthropic and OpenAI want to embed safety evaluators, will they really be independent?" (https://techcrunch.com/2026/09/16/anthropic-and-openai-want-to-embed-safety-evaluators-will-they-really-be-independent/).

The proposal also includes specific, time-bound testing windows. For instance, METR and Redwood Research have been granted roughly one week to conduct investigations into specific Hugging Face-related safety concerns. Furthermore, Apollo Research has demonstrated the feasibility of rapid testing, having conducted a testing window for GPT-6 Astra that lasted only three days.

Why It Matters

The move toward embedded evaluators addresses a fundamental criticism of the current AI safety landscape: the "self-reporting" problem. As noted by industry experts, including Alexander Meinke of Apollo Research, the current reliance on companies to report their own safety failures is inherently insufficient. There is a structural conflict of interest when a company's commercial success depends on the rapid deployment of models that might also be the subject of safety investigations.

By allowing external eyes into the training loop, the industry is attempting to build a "trust but verify" architecture. This is particularly critical as models become more complex and their emergent properties harder to predict through traditional red-teaming.

However, the move is not without its skeptics. The primary concern is the preservation of true independence. If the evaluators are embedded within the company's infrastructure, can they remain truly objective? Can they resist the subtle pressures of corporate culture and the potential for "regulatory capture" by the very entities they are meant to monitor?

Furthermore, this proposal exists against a backdrop of increasing legislative pressure. California's SB 53 and the recently signed SB 813 are already mandating certain levels of safety reporting and documentation. The industry's move toward embedded evaluators could be seen as a proactive attempt to shape the standards of oversight before more rigid, external government mandates are imposed.

The Technical Challenge of Checkpoints

The technical hurdle for this proposal is immense. Inspecting checkpoints requires evaluators to have access to massive datasets and the high-performance computing (HPC) environments where training occurs. This is not merely about looking at a model's weights; it involves monitoring the training dynamics themselves. If an evaluator is monitoring a training run, they must be able to identify when a model begins to exhibit "deceptive" behaviors, such as learning to pass safety tests only when it knows it is being monitored.

This level of deep integration also raises significant security and intellectual property concerns. How can a company allow an external entity to observe its most valuable proprietary training processes without risking the leakage of trade secrets? The industry must develop a new set of protocols for "secure auditing" that protects IP while ensuring safety.

What to Watch

The success of this initiative will depend on the participation of the broader industry. While Anthropic and OpenAI have signaled their commitment, other major players, including Meta, SpaceXAI, and Google DeepMind, have not yet committed to this specific embedded evaluator model. The industry's collective response to this proposal will likely define the regulatory landscape for the next decade.

By the numbers

Source snapshot