Today in AI — 5 May 2026

The model labs have stopped pretending they're just API companies. Both OpenAI and Anthropic announced PE-backed services ventures this week, each designed to embed their engineers inside other people's businesses. Meanwhile, the security consequences of deploying AI everywhere are getting real fast enough that the same labs are scrambling to build defences.

The consulting turn

OpenAI closed a $10 billion joint venture with 18 private equity firms; Anthropic announced a $1.5 billion equivalent backed by Blackstone, Goldman, and others. Both send embedded engineers directly inside portfolio companies. This is Accenture's playbook rebuilt around model access as the lock-in mechanism. The enterprise pitch has shifted from "here's an API key" to "here's a team that won't leave."

• OpenAI finalizes 'The Deployment Company,' a $10 billion enterprise AI venture backed by TPG and 18 PE firms — The Next Web
• Anthropic partners with Blackstone, Goldman Sachs, and Hellman & Friedman to launch $1.5 billion enterprise AI services firm — CNBC

The security arms race

Mandiant reports that mean time-to-exploit has gone effectively negative: 28.3% of CVEs exploited within 24 hours. Both labs are racing to ship security products in response. Anthropic launched Claude Security with Opus 4.7; OpenAI restricted its best cyber model to vetted defenders, weeks after criticising Anthropic for the same approach. Palo Alto is acquiring Portkey to secure the AI agents everyone else is deploying.

• AI-assisted attacks now routinely exploit vulnerabilities before patches arrive — The Hacker News
• Anthropic launches Claude Security in public beta with Opus 4.7 vulnerability scanning — Help Net Security
• OpenAI restricts GPT-5.5-Cyber to vetted defenders despite criticizing Anthropic for similar limits — The Register
• Palo Alto Networks to acquire Portkey to secure the rise of AI agents — Palo Alto Networks

The jobs argument, again

Jensen Huang called out CEOs predicting mass layoffs, specifically Dario Amodei's 50% figure, as having a "god complex." His evidence: radiology, where AI was supposed to eliminate jobs a decade ago. Meanwhile, Mayo Clinic published a study showing AI detecting pancreatic cancer up to three years before clinical diagnosis. Augmentation keeps winning over replacement, even when the capability gap is wide.

• Jensen Huang says CEOs predicting mass AI job losses have a 'god complex' — Fortune
• Mayo Clinic AI detects pancreatic cancer up to three years before clinical diagnosis in landmark study — Mayo Clinic News Network

Infrastructure and pricing shifts

Cerebras kicks off its IPO roadshow at a $26.6 billion valuation, the first pure-play AI chip company to go public. GitHub Copilot moves to per-token billing in June. Uber is turning 7 million drivers into a sensor grid for autonomous vehicle companies. Each story is a different facet of the same thing: AI infrastructure is being metered, monetised, and built out in physical space.

• Cerebras launches IPO roadshow targeting $26.6 billion valuation and $3.5 billion raise — CNBC
• GitHub Copilot switches to token-based billing starting June 1, replacing premium requests — The Decoder
• Uber turns 7 million drivers into a real-time AI sensor grid with Project Sentinel — TechCrunch

The goblin postmortem

OpenAI published an explanation of how a reward signal applied to just 2.5% of responses caused a 175% spike in goblin references across the entire model. A useful cautionary tale for anyone fine-tuning models: small reinforcement signals compound across training generations in ways that are hard to predict and harder to undo.

• OpenAI reveals how a harmless 'nerdy' reward signal turned ChatGPT into a goblin factory — OpenAI

The model labs are becoming systems integrators with billion-dollar services arms, while simultaneously racing to secure the systems they're integrating — keep an eye on whether "embedded engineer" becomes the new vendor lock-in.

---