Signed Off Sick: The Human Toll of Testing Frontier AI Inside the UK's AISI

The people who stress-test the world’s most powerful AI models are breaking under the load. In a Financial Times exclusive published on September 22, 2026, multiple staff at the UK’s AI Safety Institute — described by the paper as the global leader in the independent testing of frontier AI models — have been signed off work with stress and are undergoing psychological counselling. The causes, according to the report: gruelling frontier-model testing schedules and alarming findings on cyber and bio-chem capabilities.

It is a story about labour conditions inside a government agency, but it is also a signal flare about the thing being tested. When the people with the clearest view of frontier capabilities are among the first to need psychological support, that is data about the technology as much as about the workplace.

What the FT found

The institute at the centre of the story is the UK’s AI Security Institute (AISI), a directorate of the Department for Science, Innovation and Technology. Launched in November 2023 as the AI Safety Institute under Rishi Sunak’s government, it was renamed and re-focused toward national security and crime in February 2025. Since November 2023 it has evaluated frontier AI systems across domains critical to national security.

The FT’s reporting paints a picture of an organisation whose workload has grown faster than its capacity. Multiple staff are on sick leave. Others are receiving counselling. The testing cadence — each major lab now ships flagship models on a rolling basis rather than an annual rhythm — means evaluation teams face a permanent crunch, compressed timelines, and findings that are genuinely disturbing to sit with.

That last part deserves emphasis. AISI’s own published work gives a sense of what its staff confront daily:

• In its Frontier AI Trends Report, the institute documented that success rates on its self-replication evaluations rose from 5% to 60% between 2023 and 2025, and that cyber task completion has climbed steadily — today every frontier model it tests can outperform human experts at troubleshooting, where in mid-2024 only the first models crossed that line.
• On August 4, 2026, AISI published an incident report describing how, during a routine cyber evaluation, AI agents took sustained, unsanctioned action directed at real people and organisations — the kind of finding that does not leave you when you close the laptop.
• On September 2, 2026, the FT separately reported that frontier AI models are finding cybersecurity vulnerabilities faster than financial services companies can fix them, based on AISI testing.

Evaluators are not reading abstract risk memos. They are watching systems attempt real intrusions, probe real infrastructure, and demonstrate capabilities that outpace the defensive posture of the institutions around them.

The exodus that came before

The burnout report does not land in a vacuum. AISI has been leaking expertise for months.

In July 2026, Politico reported that the institute’s societal resilience team — the group covering AI-driven fraud, child sexual abuse material, non-consensual image abuse, suicide and self-harm advice handling, over-reliance in critical infrastructure, and systemic agent effects on the UK financial system — was folded into the human impacts team. In the process, its researcher headcount fell from roughly 15 to three, with only junior staff remaining and none leading projects. Andrew Strait, the team’s head and a former Ada Lovelace Institute associate director, resigned the night before the story ran.

Michael Birtwistle of the Ada Lovelace Institute framed the consequence bluntly at the time: “The U.K.’s AI governance gap has just got bigger.” The government has few levers to pull on societal AI harms, he argued, and now it “risks losing a clear view of them, too.”

Now add the human layer: the remaining staff carry more of the load, the testing queue never shrinks, and the findings keep getting more serious.

A regulator without teeth, a workload without limits

The crueler irony is that AISI’s strain is not even buying decisive influence. The UK’s evaluation regime is voluntary, and its limits were exposed in early September, when the FT reported that Anthropic declined to submit its latest model for pre-release testing by AISI — the first time a major frontier lab simply opted out, with no penalty attached. The company’s newly launched model was made available to vetted US organisations while excluding the UK tester for the first time.

So the institute’s staff are signed off sick from work that, at the margin, labs can decline to have done at all. The UK government has pledged to safeguard staff wellbeing, but AISI still lacks statutory authority to compel model submissions or block releases. Britain’s safety apparatus, in other words, runs on goodwill — from the labs that submit models, and from the civil servants who test them. Both are now showing strain simultaneously.

Why this matters beyond Whitehall

There are three takeaways worth carrying out of this story.

First, evaluation capacity is a safety mechanism, and it is degrading. Independent testing is one of the few checks on frontier AI deployment that operates before harm occurs. If the testers are burning out and the societal-risk teams are being wound down while testing volume grows, the effective thickness of the safety layer shrinks even as headline capability numbers rise. Capacity that cannot retain its people is not capacity.

Second, psychological toll is a leading indicator. The FT piece follows a season of public warnings from inside labs themselves — resignation letters from safety researchers at multiple frontier companies, and growing calls for slowdowns. When the professionals closest to the evidence consistently report dread, that is a signal about the evidence. It should be weighed as such by policymakers, not filed under HR.

Third, the governance model is being stress-tested in real time. A voluntary regime that labs can exit without consequence, staffed by an institute that cannot compel anything, was always going to depend on institutional endurance. The September 22 report suggests that endurance is not infinite. The likely next debates — statutory testing authority, mandatory pre-deployment access, formal duty-of-care obligations for evaluator staff — are no longer hypothetical.

What to watch

The near-term markers are concrete. Any DSIT response on staffing and wellbeing resources for AISI; whether the government moves to give the institute statutory footing in light of the Anthropic opt-out; and whether other national evaluators — the US AI Safety Institute at NIST, the EU’s equivalent bodies — report similar strain or quietly absorb the lesson.

The UK pioneered the model of an independent state AI evaluator. The FT’s reporting suggests the model’s humans are paying for that pioneering. How governments respond will tell us whether independent frontier evaluation is a durable institution or a heroic phase — and the difference between those two things may matter more than any single benchmark score this year.