The model they won't release

Anthropic released Claude Mythos Preview exclusively to twelve partner organisations through Project Glasswing on 8 April 2026, backed by $100 million in model usage credits ¹. The model autonomously identified thousands of zero-day vulnerabilities across every major operating system and browser, including a 27-year-old OpenBSD flaw that had survived five million automated tests. On the CyberGym benchmark it scored 83.1% on vulnerability reproduction, compared with 66.6% for Anthropic's previous top model.

Anthropic has explicitly stated it will not release Mythos to the public. The twelve Glasswing partners include AWS, Apple, Google, Microsoft, CrowdStrike, Palo Alto Networks, and JPMorgan. Goldman Sachs, another partner, published displacement research the same week showing AI substitutes 25,000 jobs per month , placing these institutions on both sides of the AI capability and labour displacement story.

A Tom's Hardware review challenged the marketing: the "thousands" claim rested on only 198 manual reviews, and many flagged flaws were in outdated software ². The Bessent-Powell emergency meeting suggests federal regulators took the risk seriously regardless.

Goldman Sachs published research on 6 April calculating that AI substitutes 25,000 US jobs per month and creates roughly 9,000 through augmentation, a net loss of 16,000 positions monthly ¹. Over twelve months that implies approximately 300,000 actual substitutions, against the 107,094 cumulative AI-attributed cuts that Challenger, Gray & Christmas has counted since 2023 . The gap is roughly three to one: for every job cut that appears in the public tally, Goldman's model suggests two more disappear without a press release.

Goldman's bottom-up model resolves a contradiction this topic has tracked since Update #3. The Atlanta Fed projected 502,000 AI-attributed cuts for 2026 , while the NBER found 90% of firms report zero employment impact . Goldman explains the gap: most displacement runs through attrition, contract non-renewal, and restructured job descriptions rather than announced layoffs. Workers vanish from roles that are never re-posted.

The mechanism falls hardest on entry-level positions. A study of 62 million resumes found AI-adopting firms cut entry-level postings by 15% while senior roles held flat ; the Dallas Fed confirmed the losses concentrate among workers under 25 through collapsed job-finding rates, not firing . Fortune and Columbia University research showed 75% of displaced Americans never file for unemployment insurance , meaning three quarters of the newly unemployed are invisible to the claims system policymakers depend on.

Goldman Sachs' 40-year historical analysis, published 6 April, found that workers aged 25 to 35 who are displaced early in their careers face real earnings growing 10 percentage points less than never-displaced workers over the following decade ¹. Gen Z recovers faster due to occupational mobility and AI literacy, but the immediate unemployment gap is widening.

The scarring dimension compounds the pipeline blockage documented elsewhere. A study of 62 million resumes found AI-adopting firms cut entry-level postings by 15% while senior roles held flat . The Dallas Fed confirmed the losses concentrate among workers under 25 through collapsed job-finding rates . Workers who lose positions at the start of their careers face a decade of depressed earnings; workers who cannot find entry-level positions at all face an even longer shadow.

Law school applications surged 33% year-on-year and MBA applications rose 7%, mirroring the 2008 recession's credential flight ². But ChatGPT already passes the bar exam. Gen Z workers fleeing entry-level AI disruption are accumulating $200,000 or more in debt for credentials in professions that AI is simultaneously transforming.

Challenger, Gray & Christmas confirmed cumulative AI-attributed US job cuts crossed 107,094 in April 2026 ¹. In March, AI led all stated reasons for US layoffs for the first time on record, with 15,341 AI-attributed cuts in a single month . The attribution share jumped from roughly 10% in February to 25% in March.

Oracle's 30,000-person cut likely inflated the month, but even excluding it, the trend line is accelerating: full-year 2025 saw 5% AI attribution; Q1 2026 averaged 13%. Goldman's bottom-up model implies the headline Challenger figure covers only one-third of actual substitutions occurring through attrition and non-renewal. For every cut that appears in the public tally, two more disappear through roles that are quietly restructured or never re-posted.

The acceleration is sharpest in technology. Tech sector Q1 2026 cuts reached 52,050, up 40% year-on-year. A Challenger executive noted that AI replacing coding functions in technology companies is where "the actual role replacement is visible."

Oracle WARN Act filings remained incomplete as of early April 2026 ¹. Washington state filed 491 positions; Missouri filed 539. Massachusetts produced no filing despite Oracle's Burlington offices. At up to 30,000 cuts , total US WARN filings cover less than 4% of the affected workforce. Law firms are investigating potential violations.

The Massachusetts gap follows the pattern the New York WARN Act study documented: zero AI attributions from 162 companies covering 28,300 workers in the law's first year . Oracle's 12,000 India staff were terminated by a 6am email , and none will appear in US BLS payroll data or any WARN Act filing. The company has demonstrated a template for circumventing US disclosure law: concentrate cuts offshore, use email terminations without HR contact, and file minimally where state laws require it.

The 60-day clock from Oracle's 31 March cuts expires in late May. If no Massachusetts filing appears, the precedent is set.

Tom's Hardware published a critical review of Anthropic's Mythos claims on 9 April, noting that the "thousands of zero-days" assertion rested on only 198 manual reviews ¹. Many of the flagged vulnerabilities were in outdated software no longer in active use. The gap between Anthropic's marketing language and the verified sample is wide enough to warrant caution.

The Bessent-Powell emergency meeting at Treasury headquarters proceeded regardless of this scrutiny. Challenger data confirmed AI-attributed cuts crossed 107,094 the same month , suggesting federal regulators assessed the systemic risk of AI broadly, beyond Mythos's specific claims. Whether Mythos found hundreds or thousands of exploitable flaws, the CyberGym benchmark score of 83.1% versus 66.6% for its predecessor represents a measurable capability jump that the twelve Glasswing partners will deploy in production environments.

China's Ministry of Human Resources and Social Security recognised 42 new AI-related occupations in April 2026, each projected to require 300,000 to 500,000 workers ¹. The ministry is preparing a dedicated AI employment policy covering 12.7 million graduates, including job-retention rebates, social security subsidies, and five targeted training programmes.

The contrast with Western approaches is sharp. The EU voted to delay AI Act employment rules by 16 months , while the US has produced no federal AI workforce legislation with a viable path. China previously positioned AI as an employment engine in its five-year plan ; recognising 42 new occupations formalises the strategy. China deploys the state as a workforce intermediary; the US and EU treat AI displacement as a market phenomenon to be measured rather than managed.