Skip to content
You can now search across every topic, entity and event.What's new
AI: Jobs, Power & Money
10APR

The model they won't release

3 min read
16:54UTC

Anthropic's Claude Mythos Preview, a frontier AI model deliberately withheld from public deployment, triggered the first emergency convening of bank CEOs by the US Treasury and Federal Reserve over a single AI capability.

Key takeaway

Regulators moved in 48 hours on an AI cybersecurity threat but not three years of displacement data.

This briefing mapped
Loading map…
Regulatory
Competitive
Economic
Legal
Domestic

The first emergency meeting convened by US regulators over a single AI model's capabilities drew five Wall Street CEOs to Treasury headquarters.

Sources profile:This story draws on centre-left-leaning sources from United States
United States

Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell summoned the CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs to Treasury headquarters on 8 April 2026 to discuss Claude Mythos Preview. It is the 1st recorded federal emergency meeting convened over a frontier AI model's capabilities.

Financial sector AI adoption grew 127% year-on-year as of 3 April. Over 107,000 AI-attributed job cuts across 3 years produced no equivalent government response. 

Sources:Bloomberg
Briefing analysis

The pattern of institutional emergency response for financial stability paired with institutional inertia on workforce protection has a direct precedent. In September 2008, the US Treasury and Federal Reserve convened bank CEOs within 72 hours of Lehman Brothers' collapse, producing a $700 billion rescue package (TARP) within two weeks. The foreclosure crisis that displaced 10 million American homeowners from 2007 to 2012 produced the Home Affordable Modification Program only in March 2009, six months after the bank rescue, and ultimately reached fewer than a third of eligible borrowers.

The structural parallel is precise. Financial system instability triggers emergency federal convening with named officials, specific participants, and rapid institutional response. Household instability triggers programmes that are slower, less well-funded, and reach fewer of the affected. The Bessent-Powell meeting over Mythos follows this template: a frontier AI capability that threatens bank infrastructure produced an emergency convening within 48 hours. Three years of AI displacement crossing 100,000 attributed cuts (and likely far more) has produced bipartisan letters, proposed bills that die in committee, and data collection requests to agencies with no mandate to act.

Anthropic's most capable model scored 83.1% on vulnerability reproduction but will not be released publicly, going instead to twelve partners through a $100 million restricted programme.

Anthropic released Claude Mythos Preview on 8 April 2026 exclusively to 12 partners through Project Glasswing, backed by $100 million in model usage credits. It scored 83.1% on the CyberGym benchmark versus 66.6% for Anthropic's previous top model, and found a 27-year-old OpenBSD flaw that had survived 5 million automated tests.

Anthropic will not release Mythos publicly. It is the 1st time a frontier lab has withheld a model from public release on dual-use safety grounds. 

Sources:Anthropic

A bottom-up displacement model from Goldman Sachs calculates AI is eliminating three times more jobs per month than appear in any official tally.

Sources profile:This story draws on mixed-leaning sources from United States
United States
LeftRight

Goldman Sachs published research on 6 April 2026 calculating that AI substitutes 25,000 US jobs per month and creates roughly 9,000 through augmentation, a net loss of 16,000 monthly. Over 12 months that implies 300,000 actual substitutions, against the 107,094 cuts Challenger, Gray & Christmas has counted since 2023.

The gap is roughly 3 to 1. Most displacement runs through attrition and contract non-renewal rather than announced layoffs, making it invisible to the public tally. 

Goldman's 40-year historical analysis reveals that workers displaced between 25 and 35 never fully recover their earnings trajectory.

Sources profile:This story draws on mixed-leaning sources from United States
United States
LeftRight

Goldman Sachs' 40-year analysis, published 6 April 2026, found workers aged 25 to 35 displaced early in their careers earn 10 percentage points less in real terms than peers over the following decade. Gen Z recovers faster through AI literacy and mobility, but the immediate unemployment gap is widening.

Law school applications surged 33% year-on-year as workers flee entry-level disruption. ChatGPT already passes the bar exam, so workers accumulating $200,000 in debt for legal credentials face disruption a 2nd time. 

AI led all stated reasons for US job cuts in March for the first time on record, pushing the cumulative tally past 100,000.

Challenger, Gray & Christmas confirmed cumulative AI-attributed US job cuts crossed 107,094 in April 2026. AI led all stated layoff reasons in March for the first time, with 15,341 cuts in that month alone. The attribution share jumped from roughly 10% in February to 25% in March.

Oracle's 30,000-person cut likely inflated March's figure, but the trend is real: full-year 2025 saw 5% AI attribution; Q1 2026 averaged 13%. Goldman estimates announced cuts cover only one-third of actual AI substitutions. 

Fewer than 1,100 of Oracle's up to 30,000 cut positions appear in US disclosure filings, with Massachusetts producing no filing at all.

Sources profile:This story draws on neutral-leaning sources

Oracle's WARN Act filings remained incomplete as of early April 2026: Washington state filed 491 positions, Missouri 539, and Massachusetts filed nothing despite Oracle's Burlington offices. Total US filings cover under 4% of Oracle's up-to-30,000 cuts. Law firms have opened potential violation investigations.

Oracle's 12,000 India staff, terminated by a 6am email , produce no US disclosure entry. The 60-day clock from Oracle's 31 March cuts expires in late May. 

A technical review found Anthropic's marketing relied on 198 manual reviews to support claims of thousands of severe vulnerabilities.

Sources profile:This story draws on neutral-leaning sources

Tom's Hardware published a critique of Anthropic's Mythos on 9 April, noting the "thousands of zero-days" claim rested on only 198 manual reviews and that many flagged vulnerabilities were in outdated software no longer in active use.

Anthropic's marketing language overstates the verified 198-review sample. The CyberGym benchmark score of 83.1% versus 66.6% for its predecessor is independently verifiable, and the Bessent-Powell emergency meeting at Treasury suggests regulators assessed the model's capability trajectory regardless. 

Beijing is building the workforce pipeline while the US and EU debate whether to measure displacement at all.

Sources profile:This story draws on neutral-leaning sources

China's Ministry of Human Resources and Social Security recognised 42 new AI-related occupations in April 2026. Each category is projected to require 300,000 to 500,000 workers, and the ministry is preparing a dedicated AI employment policy covering 12.7 million graduates.

The EU delayed its workplace AI rules by 16 months; the US has no federal AI workforce legislation on a viable path. China is building state-directed workforce pipelines while Western governments debate measurement. 

Closing comments

Escalating on the capability side; stalled on the protection side. The CyberGym benchmark jump from 66.6% to 83.1% in a single model generation, combined with the Treasury-Fed emergency response, signals the AI capability curve is accelerating faster than institutional frameworks anticipated. Financial sector AI adoption growing 127% year-on-year confirms the deployment pace is matching the capability pace. On workforce protection, the trend is flat to negative. The EU delayed its binding workplace AI rules by 16 months. The Sanders-AOC moratorium was killed by the Democratic caucus. The only surviving federal action is a bipartisan data-collection letter with no enforcement mechanism. The DOL apprenticeship AI training initiative is the most concrete positive action but addresses a narrow cohort. The 55% leader-regret finding and Klarna's public reversal introduce a market correction signal that policy has not produced. If the pattern generalises, some over-correction will reverse without intervention. The Goldman scarring data suggests the workers already displaced during the correction period will carry the earnings loss regardless of whether the macro trend moderates.

Different Perspectives
US Treasury and Federal Reserve
US Treasury and Federal Reserve
Bessent and Powell convened five Wall Street bank CEOs within 48 hours of learning of Mythos's cybersecurity capabilities, while the same banks now hold privileged Glasswing access and Goldman simultaneously publishes data showing AI eliminates a net 16,000 US jobs per month. The regulatory urgency applied to AI capability has no parallel in three years of displacement data.
Klarna and European fintech
Klarna and European fintech
CEO Sebastian Siemiatkowski publicly admitted Klarna's AI customer service cuts went too far and announced rehiring under a hybrid model, the first major company to reverse an AI-driven workforce reduction and confirm what 55% of surveyed leaders already reported: the cuts were wrong.
China
China
China's Ministry of Human Resources and Social Security recognised 42 new AI occupations each requiring up to 500,000 workers, and is preparing an employment policy covering 12.7 million graduates, treating the transition as a state workforce management problem rather than a market phenomenon.
UK workers and policymakers
UK workers and policymakers
UK firms suffered net AI-driven job losses of 8% over the past year, double the international average despite identical productivity gains to US peers, while the OBR has modelled 500,000 additional unemployed in a worst-case scenario the Bank of England plans to stress-test in its banking assessments.
EU regulators
EU regulators
The EU voted 101 to 9 to delay AI Act workplace rules by 16 months, stripped employer AI literacy obligations from Parliament's position, and faces a final trilogue on 28 April where the binding worker rights language remains contested, leaving EU workers without enforceable protections as AI deployment accelerates.
Researchers and academic economists
Researchers and academic economists
The Hamilton Project and PIIE finding that entry-level hiring declines preceded ChatGPT's launch introduces a fundamental causal question: if rising interest rates drove the initial slowdown, AI is being used as ex-post cover for cuts with a different origin, complicating both the displacement count and the policy response.