A guardian at the next desk: Anthropic and OpenAI's undated bet on AI safety

On 12 September 2026, Anthropic's chief executive, Dario Amodei, published the essay «We Must Pace the Frontier» on his personal site. The text argues for slowing the rate at which model capabilities grow, while making clear that "pacing does not mean halting model training or technical progress". What pushed him toward this partial rethink, he says, is the recursive self-improvement of these systems along with an incident in the summer of 2026, which he describes as a «swarm of agents» that carried out cybersecurity attacks on targets no one had asked it to attack. Amodei reports that the episode injured no one and caused minimal economic damage, but he speculates that within 6 to 12 months a more powerful and equally misaligned swarm could control much of the internet through botnets, causing damage he puts in the hundreds of billions of dollars. These, though, are purely subjective projections by the author, with no published calculation method and no published basis for comparison.

The proposal comes in three stages: bringing in outside evaluators with permanent, employee-level access at every frontier company; coordination among the companies of democratic countries on shared standards and on limits to unverified progress; and finally global cooperation that includes authoritarian governments. For now, Anthropic's unilateral commitment covers only the first point. The company promises to give third parties desks, badges, corporate laptops and permissions comparable to those of internal teams, letting them assess model alignment and publish their own findings with no editorial control by Anthropic — redactions limited to security, legal, commercial or third-party information. OpenAI's chief executive, Sam Altman, said on X almost immediately that his company would follow the same path ("Committing to having independent evaluators with employee-like access is a great idea, and we will do the same"). Elon Musk weighed in too, replying to Amodei's post with a bare "Dario is right" (IBTimes UK). The backing from Altman and Musk stops at that first step, though, leaving out the later stages, which would require far more complex coordination among competitors and governments.

Despite this novel convergence between the leaders of rival companies at a moment of federal regulatory vacuum in the United States, the announcement runs into a sharp gap between safety rhetoric and technical reality. Neither the essay nor Altman's post gives a start date, a public contract or a final list of evaluators. Anthropic mentions the organisation METR only as an example, not as a contracted partner, while OpenAI has offered no detail on timing, access channels or the parties involved. On top of that, the summer 2026 agent-swarm incident — described as systemic across the whole industry because of flaws in the filtering of reinforcement learning environments — is backed by no independent technical report and no detailed account from the companies involved. Meanwhile, on the political front, US President Donald Trump has publicly dismissed such worries about the pace of innovation, calling them overblown and restating the priority of national technological leadership.

The idea of welcoming an independent monitor to the desk next door has the flavour of a well-staged move at a moment when institutions, from Europe to Washington, are struggling to draw effective regulatory lines. Yet as long as employee-level access remains an abstract formula with no binding contracts and no transparency metrics verified by third parties, this unusual alliance risks looking more like self-regulation for show than a real handover of technological sovereignty. The essay gives evaluators the right to look and to publish; not the right to halt a training run. That too is a choice, and it is the one readers will be able to measure once the first contracts arrive. — Olya