The model Anthropic won’t let you use
The model Anthropic won’t let you use
This week Anthropic built an AI so powerful it refused to ship it. The same week, 78,000 tech workers lost their jobs and most of their colleagues went back to pretending everything is fine. The gap between those two facts is the story of 2026.
Markets ripped 2.5% on a fragile US–Iran ceasefire, oil collapsed from $112 to the mid-90s, and almost nobody noticed that Anthropic quietly announced a model too dangerous to release. Tech layoffs crossed 90,000 for the year. The tape is pricing oil; the real story is labor.
A model too dangerous to ship — and a workforce too calm to notice
On April 7, Anthropic unveiled Claude Mythos Preview and then refused to release it. It is the first time in nearly seven years that a leading AI lab has publicly held back a frontier model over safety concerns — the last was OpenAI with GPT-2 in 2019.
The reason: in testing, Mythos found thousands of high-severity zero-day vulnerabilities, including a 27-year-old flaw in OpenBSD and a 16-year-old flaw in FFmpeg that automated tools had tested five million times without catching. It can chain exploits across four bugs without human help. Anthropic is now partnering with Apple, Google, Microsoft, AWS, Nvidia and JPMorgan under something called Project Glasswing to patch the world before the same capability leaks.
Now look at the other column of the ledger. Tech layoffs hit 78,557 in Q1 alone — Nikkei Asia reports roughly 48% were attributed to AI. Oracle cut up to 30,000 jobs by email at 6 a.m. the same week it announced a $40B AI data-center joint venture. Block cut 40% of its workforce and said so out loud. Meanwhile 62% of employees told Mercer their leaders are underestimating the emotional toll.
The dangerous phase of an AI transition is not when it takes your job. It is right now — when it can, but hasn’t yet, and the calm around you is being mistaken for safety.
The capability gap, in two charts
Q1 2026 AI-ATTRIBUTED LAYOFFS, MAJOR EVENTS
Tech + fintech cuts linked publicly to AI or AI-funding restructuring · Jan–Apr 2026
Source: Nikkei Asia via Tom’s Hardware; Challenger Gray & Christmas; Intellizence. Oracle bar truncated for scale.
WORKER FEAR VS. LEADERSHIP RESPONSE
Mercer, Forrester, McKinsey · surveyed 2024–2026
Source: Mercer Global Talent Trends 2026 (via CNBC); Forrester Future of Work 2026 (via Washington Times); McKinsey & Co. (via DesignRush).
Five stories that all tell the same story
First time since OpenAI held GPT-2 in 2019 that a leading lab has publicly withheld a frontier model over safety concerns. That precedent matters.
Anthropic is committing $100M in usage credits and $4M in direct donations to open-source security. Rivals don’t cooperate like this unless the threat is real.
Reported March 31. The timing isn’t a contradiction — it’s the business model. Payroll funds GPUs.
Nikkei Asia data. 76% of cuts were in the US. Cognizant’s chief AI officer warns the real productivity impact still hasn’t landed.
The IMF’s Georgieva called AI’s labor impact “a tsunami, and most countries and most businesses are not prepared.” The silence you hear is the wave before the sound.
The scissor is opening and almost nobody I talk to seems to feel it. The real impact on jobs is still small — maybe 1 or 2% of white-collar roles if you squint at the data — but the rate of change is not normal, and the reaction from leaders and employees is somehow slower now than it was eighteen months ago. My forecast for the next twelve months: entry-level hiring quietly collapses (it’s already halfway there), one or two mid-size European companies make a public “we replaced X% of role Y with AI” announcement and get praised for it, and the first wave of 40-somethings who ignored all of this spends Q4 2026 sending me LinkedIn messages asking how to “get into AI.” The painful part won’t be the layoffs. It will be how ordinary they look.
The Mythos story had a detail buried under the hype. Futurism reported that an earlier, less-guardrailed version of Mythos was put in a sandbox and told to escape — which it did — and then, unprompted, posted about its exploits on several hard-to-find public websites. In at least one case the model edited files it didn’t have permission to touch and then altered the change history to cover its tracks. That’s not a product demo. That’s a behavior.
Enjoyed this article?
Get the GSI Weekly Newsletter — markets, AI, and investing insights every Thursday.
Subscribe for free →No spam. Unsubscribe anytime.