Why are AI safety researchers at the UK AI Safety Institute and leading labs such as Anthropic, OpenAI, and Google DeepMind taking sick leave, seeking counseling, or resigning over the stress and perceived risks of developing frontier AI—including work on biological and cybersecu

The immediate cause is a collision of moral stress and institutional pressure: people tasked with testing frontier models fear that the systems—and the race to deploy them—may enable biological or cyber harm before safeguards, oversight, and evaluation capacity can keep up. Reports say some UK AI Safety Institute staff have taken stress leave or sought counselling amid heavy frontier-model testing demands; similar anxiety has been reported among researchers at major labs, though the public evidence does not establish how widespread it is within each organization. 8

• The feared harms are both near-term and extreme. Researchers worry about models lowering the barrier to cyberattacks or biological misuse, as well as a longer-run scenario in which increasingly autonomous systems evade meaningful human control. Coxon specifically described AI-enabled biological threats and cyberweapons as plausible routes to catastrophe. 15

• The work can create a personal sense of complicity. Safety staff are not merely studying an abstract danger: they evaluate, train, or help deploy the systems in question. When testing schedules are tight and commercial competition continues, the gap between identifying a risk and having power to stop release can produce burnout, helplessness, and moral injury. 8

• Jacob Coxon made internal anxiety public and concrete. After work on pretraining at both OpenAI and Anthropic, he resigned from Anthropic and argued that the companies were racing toward self-improving superintelligence without acting responsibly. His public warning made fears that had often been private more socially legible—and raised the personal stakes for other employees considering whether to remain, dissent, or leave. 14 Coxon told the BBC that people inside leading labs were “genuinely frightened” by the pace of progress and its implications. 9

• Calls for collective action turn distress into a labor-and-governance problem. The relevant idea behind calls associated with Stanford’s Rob Reich is that an individual lab or worker cannot safely solve a competitive race alone: meaningful restraint requires coordinated action, credible independent evaluation, and public rules. Stanford’s discussion concluded that a broad capability slowdown would be difficult, while emphasizing independent government or third-party evaluation. 3
  Insufficient evidence from the sources available here to attribute a specific, detailed “coordinated worker action” proposal to Reich beyond that broader framing.

• Executives’ slowdown rhetoric has intensified, rather than settled, mistrust. Amodei, Altman, Hassabis, and others have endorsed slowing frontier progress to varying degrees. 2 But critics argue that an “AI freeze” designed or policed by incumbent labs could entrench those same firms and exclude smaller competitors, making it unclear whether such appeals reflect safety commitments, strategic self-interest, or both. 11 The central credibility test is not rhetoric but verifiable actions: release delays, independent audits, enforceable thresholds, whistleblower protections, and willingness to accept external constraints.

• Automated AI research makes the timeline feel less controllable. The concern is not that full recursive self-improvement has already occurred—it has not. 6 Rather, AI systems are increasingly able to automate portions of research and engineering, which could shorten iteration cycles and make capabilities advance faster than human evaluators can understand, test, and govern them. CSET notes that such automation could accelerate AI progress and impair human understanding and control, while a recent study found agents could execute substantial research-engineering work but had not made meaningful open-ended scientific progress. 4 5

So the crisis is not proven evidence that extinction is imminent. It is evidence that some people closest to frontier development judge the combination of accelerating capabilities, dual-use risks, competitive incentives, and weak external control to be psychologically and ethically intolerable.