Top Stories
Anthropic Withholds Its Most Powerful Model, Citing Cybersecurity Risk
Anthropic released Claude Mythos Preview on April 7 — then immediately said the public can’t have it. The model found thousands of high-severity vulnerabilities across every major operating system and web browser, including a 27-year-old remote crash bug in OpenBSD and a 16-year-old flaw in FFmpeg that automated tools had tested five million times without catching.
This isn’t a marketing stunt dressed up as caution. Mythos can chain together three, four, or five vulnerabilities that individually do nothing but in sequence produce sophisticated exploits. Anthropic says the model is now competitive with the best human security researchers, and it has privately warned government officials that Mythos makes large-scale cyberattacks significantly more likely this year.
Instead of a public release, Anthropic launched Project Glasswing: over 50 organizations — including Microsoft, Apple, Google, Nvidia, Cisco, CrowdStrike, and JPMorganChase — get access to Mythos Preview with more than $100 million in usage credits, exclusively for defensive security work. The bet is that getting critical infrastructure patched before adversaries can replicate these capabilities is worth more than any API revenue. Whether that calculation holds depends entirely on how long the capability gap lasts. History suggests: not long.
Source: Hacker News, VentureBeat, NBC News, Anthropic
Tech Layoffs Hit 80,000 in Q1 — Nearly Half Blamed on AI
The numbers are getting harder to ignore. Almost 80,000 tech workers lost their jobs in the first quarter of 2026, with 47.9% of cuts attributed directly to AI replacing human roles. In March alone, AI was cited as the leading reason for layoffs for the first time, accounting for 25% of all stated reasons — up from 10% just one month earlier.
The biggest single event was Oracle cutting an estimated 20,000–30,000 employees on March 31. Amazon shed 16,000 corporate roles in January. Dell cut 11,000. Block eliminated 4,000 jobs — 40% of its workforce. Atlassian laid off 1,600, then immediately announced 800 new AI-focused hires, making the restructuring explicit: out with the old skill set, in with the new.
The Challenger report shows hiring plans actually rose 157% in March to nearly 33,000 positions. But the new jobs aren’t the same jobs. They’re AI engineering, ML operations, and AI safety roles. The labor market isn’t shrinking so much as it’s being rebuilt around a different set of skills, and the transition is happening faster than anyone’s retraining programs can keep up.
Source: Tom’s Hardware, Challenger, Gray & Christmas
OpenAI Preps IPO, Promises Retail Investors a Seat at the Table
OpenAI CFO Sarah Friar told CNBC the company will reserve a portion of IPO shares for retail investors — a notable break from standard tech IPO playbooks where individual investors typically get 5–10% of allocations. Friar said the company “tested the waters” with retail in its latest funding round and saw demand three times what it expected: OpenAI set out to raise $1 billion from individuals through private placements and ended up pulling in $3 billion.
The numbers backing the IPO push are substantial. OpenAI crossed $25 billion in annualized revenue at the end of February, up from $21.4 billion at year-end 2025. Enterprise now makes up 40% of revenue and is on track to match consumer by year’s end. The company closed a $122 billion funding round in March at an $840 billion post-money valuation and is reportedly targeting a $1 trillion valuation for the public offering, with a filing possible in the second half of 2026.
Friar wouldn’t commit to a timeline but said it’s “good hygiene” for an $852 billion company to “look and feel and act like a public company.” Which is corporate-speak for: the filing is coming, and probably sooner than the careful non-answers suggest.
Quick Hits
-
Perplexity’s agent pivot pays off big: ARR jumped 50% in a single month to $450 million after the company shifted from AI search to agentic products and usage-based pricing. The “Computer” agent tool, launched February 25, drove the surge. Over 100 million monthly active users, with subscriptions running $20–$200/month. Prior ARR was $305 million — the growth curve is steepening, not flattening. PYMNTS, Yahoo Finance
-
DeepSeek V4 targets late April launch on Huawei silicon: The 1-trillion-parameter MoE model will reportedly run on Huawei’s Ascend 950PR chips — the first frontier AI model that doesn’t depend on Nvidia. Only 32–37 billion parameters activate per token, keeping inference costs manageable. DeepSeek deliberately gave Huawei early access while denying it to Nvidia, signaling China’s intent to break CUDA dependence entirely. If the benchmarks hold up, the geopolitical implications are significant. GizChina, TrendForce
-
Arcee ships a 399B open-source model under Apache 2.0: Trinity-Large-Thinking, from a 30-person San Francisco startup, activates only 13B parameters per token via MoE, runs 2–3x faster than dense models of comparable capability, and costs $0.90 per million output tokens — 96% cheaper than Claude Opus 4.6. Arcee bet $20 million (half its total funding) on a single 33-day training run using 2,048 Nvidia B300 GPUs. As the major labs drift toward proprietary enterprise models, a genuinely open U.S.-made frontier model is increasingly rare. VentureBeat, MarkTechPost
-
India’s Sarvam AI closing $350M raise at $1.5B valuation: Backed by Bessemer, Nvidia, and Amazon, Sarvam is building sovereign AI models for 22 Indic languages. The Indian government subsidized 4,096 H100 GPUs through the IndiaAI Mission. Microsoft’s $17.5B India commitment is materializing with a Hyderabad data center going live this summer. Total AI venture funding for Indian companies now exceeds $2.9 billion. Outlook Business
Worth Watching
The Mythos situation deserves more scrutiny than the “too dangerous to release” framing suggests. Anthropic’s decision to restrict access is defensible on its face — if a model can find exploitable zero-days in every major OS, making it publicly available would be reckless. But the Project Glasswing structure also means Anthropic gets to position itself as the indispensable cybersecurity partner for the world’s biggest tech companies, with $100 million in credits creating deep integration that’s hard to walk away from. Good security policy and good business strategy aren’t mutually exclusive, and both appear to be operating here.
Meanwhile, the layoff numbers and the IPO numbers tell the same story from opposite directions. AI is destroying jobs fast enough to dominate the Challenger report for the first time, while simultaneously generating revenue growth that justifies trillion-dollar valuations. The question isn’t whether AI creates more value than it destroys — by most financial metrics, it clearly does. The question is who captures that value, and whether the 80,000 people who lost their jobs this quarter are anywhere in the distribution.