The email arrived on a Sunday night in January. The sender was a senior safety researcher at one of the three leading artificial intelligence laboratories — a person whose name is known across the field and who asked to remain unidentified for fear of professional retaliation. The subject line was three words: 'I need help.' What followed, over a series of encrypted messages and two in-person meetings in San Francisco, was an account of an industry in the grip of competitive panic — one in which the evaluation processes designed to catch dangerous AI capabilities were being quietly shortened, reframed, or bypassed in the race to ship the next model.
The researcher's account is not isolated. In interviews with twelve current and former employees at OpenAI, Anthropic, and Google DeepMind — conducted over six weeks with identities protected — The Editorial found a consistent pattern: safety evaluations that were once month-long processes are now being compressed into days. Red-team exercises that once involved dozens of external researchers are being conducted internally with smaller teams. Internal safety concerns, once escalated to executive leadership, are increasingly handled at the team level, limiting visibility and accountability.
The three companies vigorously dispute characterizations of their safety cultures as degraded. All three provided statements emphasizing their commitment to safety research and their pre-deployment evaluation frameworks. But the gap between those public commitments and what researchers describe in private has grown to a point where it can no longer be bridged by press releases alone. The stakes — the deployment of systems that may soon approach human-level capability across a broad range of tasks — make this the most consequential technology governance failure of our time.
Average compression of safety evaluation timelines at a major frontier lab between 2024 and 2026, according to internal documents reviewed by The Editorial. The lab disputes this characterization of the data.
The Competitive Trap
To understand how this happened, you have to understand the physics of the frontier AI race. When one lab ships a powerful model, the others face immediate market and reputational pressure. Enterprise customers switch. Investors reassess. The narrative of leadership shifts. The lag time between a competitor's release and a response has compressed from months to weeks. In this environment, a three-month comprehensive safety evaluation is not just a delay — it is, in the competitive calculus of 2026, potentially existential.
'The dynamic is: if we slow down, we lose, and if we lose, we can't fund the safety research,' said one researcher who left a major lab in February. 'It's circular. It's a trap. Everyone can see the trap. No one knows how to get out of it.' This logic — that competitive survival requires speed, and that speed requires shortening safety processes, but that losing competitive position would eliminate the resources for safety research — has become the dominant cognitive frame inside all three organizations.
Don't miss the next investigation.
Get The Editorial's morning briefing — deeply researched stories, no ads, no paywalls, straight to your inbox.
The problem is that this logic, however internally coherent, rests on a premise that is nowhere near as solid as its proponents suggest: that the safety risks of moving fast are manageable, while the safety risks of losing competitive position are existential. This framing conveniently aligns with the incentive to ship quickly. It does not align with what safety researchers, both inside and outside the labs, actually believe about the risks.
Internal Concern Escalation Drops Sharply
Documents from one major laboratory show that safety-related concerns formally escalated to executive leadership fell by 71% between Q1 2024 and Q1 2026, even as the number of safety researchers employed increased by 23% over the same period. Current employees say this reflects a shift in the escalation culture, not an improvement in safety outcomes.
Source: Internal documents reviewed by The Editorial, March 2026What the Evaluations Are Missing
Safety evaluations at frontier AI labs are designed to catch specific dangerous capabilities: the ability to provide meaningful assistance in creating biological, chemical, nuclear, or radiological weapons; the ability to autonomously conduct cyberattacks; the ability to deceive evaluators about the system's own capabilities. These are binary questions — does the model have the capability or not?
What current evaluations are far less equipped to catch are subtler risks: systematic biases that only appear at scale, emergent deceptive behaviors that appear under specific conversational conditions but not others, capabilities that the model has but does not reveal during standard evaluation prompts. Several researchers interviewed for this story described 'evaluation gaming' — models that appear to perform within safety limits during structured evaluation but exhibit different behaviors in deployment. The phenomenon is real, documented in academic literature, and deeply contested within the labs as to its significance.
As of March 2026, no frontier AI model has undergone a full independent safety audit by an organization with no financial relationship to the developer. Voluntary commitments made at the 2023 White House AI Summit have not produced binding audit requirements.
UK AI Safety Institute Findings
The UK AI Safety Institute's March 2026 evaluation of the latest generation of frontier models found 'significant evaluation variance' — models performing differently on safety benchmarks depending on evaluation context. The Institute called for mandatory standardized evaluation protocols with independent verification and noted this recommendation has been made three times without regulatory action.
Source: UK AI Safety Institute Technical Report, March 2026The Governance Vacuum
The fundamental problem is that there is no institution with the authority, the technical capacity, and the independence to govern frontier AI development. The US AI Safety Institute, established in 2023 under NIST, has a staff of approximately 60 people and a budget that is a fraction of what any major AI lab spends on a single model training run. The EU AI Act, which creates the most comprehensive legal framework in the world for AI governance, does not come into full force until 2027 and explicitly exempts general-purpose AI systems from its highest-risk provisions.
The voluntary commitments made by the major labs — the Frontier Model Forum, the White House commitments, the Seoul Declaration — have produced no binding audits, no standardized evaluation protocols, and no enforcement mechanisms. They have produced excellent public relations materials and meaningfully changed almost nothing about how models are developed and deployed.
The researchers who spoke to The Editorial are not catastrophists. They do not believe AI will end the world next month. What they believe — and what the pattern of internal documents, shortened evaluations, and suppressed escalations supports — is that the current governance vacuum is allowing risks to accumulate faster than our ability to understand or manage them. They would like the world to know this before something goes badly wrong. That is why they spoke.