🔥 Tech Intelligence Digest — May 31, 2026

Covering: May 29–31, 2026

Anthropic raises $65B at ~$965B valuation — Series H round makes it the most valuable private AI company, with an October 2026 IPO expected around $1T+ (TechCrunch, Bloomberg)
Cognition (Devin) raises $1B at $26B valuation — AI coding agent hit ~$492M ARR; 90% of Cognition's own code is written by Devin (TechCrunch)
OpenAI launches Rosalind Biodefense — GPT-Rosalind life-sciences model now available to trusted developers for pandemic preparedness (OpenAI)
Illinois passes landmark AI safety bill (SB 315) — Requires annual third-party safety audits for frontier AI labs; OpenAI and Anthropic both support it (NBC News)
Google confirms first AI-developed zero-day exploit — Criminal group used an AI model to craft a 2FA bypass vulnerability (Politico)

AI and Models

Anthropic's $965B moment. The Series H round is the largest private AI raise in history. Goldman Sachs, JPMorgan, and Morgan Stanley are lined up for an October IPO targeting $1T+. Theo (t3.gg) put out a video titled "Holy sh*t I think Anthropic is profitable now" (video) — worth watching for his take on the revenue trajectory.

OpenAI's dual push: biodefense + tax automation. Two distinct launches this week:

Rosalind Biodefense (May 29): GPT-Rosalind is a frontier life-sciences reasoning model. The program gives trusted developers and government agencies access to build pandemic preparedness tools — epidemiological modeling, outbreak detection, countermeasure development. Framed as "defensive acceleration" (OpenAI).
Self-improving tax agents with Codex (May 27): Tax AI uses production usage data as structured feedback to drive an autonomous improvement loop. Demonstrates the "self-improving agent" pattern applied to enterprise workflows (OpenAI).

OpenAI's Frontier Governance Framework (May 28) and a shared playbook for third-party evaluations (May 29) signal the company positioning itself ahead of regulation (OpenAI).

Dev Tools and Infrastructure

The coding agent wars are heating up fast.

Cursor hit a $3B annualized revenue run rate (up from $2B in February). 3,000+ enterprise customers paying $100K+/year each. SpaceX reportedly has an option to acquire Cursor at ~$60B, with Composer 2.5 trained on SpaceX infrastructure (Bloomberg).
Cognition (Devin) raised $1B at $26B. ARR grew from <$1M (Sep 2024) → $73M (mid-2025) → $492M (May 2026). Customers include Goldman Sachs, NASA, US Army, Mercedes-Benz. 89% of Cognition's own commits are Devin-authored (TechCrunch).
Canyon Code closed a $5M pre-seed (Cota Capital) for a workflow intelligence layer for AI agents (SiliconANGLE).
Replit got strategic backing from Visa, integrating its Trusted Agent Protocol.

Infrastructure spending: Big Tech's 2026 CapEx is tracking ~$715-725B. Amazon leads at ~$200B, Microsoft and Alphabet at ~$190B each, Meta at $125-145B (Intellectia).

Industry and Big Tech

Illinois passes SB 315 — the most comprehensive state-level AI safety law in the US. Key requirements:

Annual independent third-party safety audits for frontier AI labs
Mandatory incident reporting for safety breaches
Whistleblower protections for AI company employees
Covers "catastrophic risks" including CBRN misuse and critical infrastructure disruption
Passed House 110-0, Senate 52-5. Both OpenAI and Anthropic publicly supported it.

This matters because federal AI regulation has stalled. Illinois is setting the template other states will copy. (NBC News, Ars Technica)

OpenAI told investors it plans $600B in compute spending by 2030. The AI infrastructure race is scaling to jaw-dropping levels (DC News Now).

Security

Google confirms first real-world AI-developed zero-day exploit. A criminal group used an AI model to discover and weaponize a logic vulnerability that bypassed 2FA in an open-source web admin tool. The exploit was a Python script with "educational" docstrings and a hallucinated CVSS score — classic AI-generated code fingerprints. Google patched it before mass exploitation. (Politico, NYT)

BadHost (CVE-2026-48710) — Critical vulnerability in an open-source package used by millions of AI agents. Allows trivial exploitation against unpatched deployments. AI infrastructure components themselves are becoming high-value targets. (Ars Technica)

The patch window is dead. Mandiant's M-Trends 2026 data: 28.3% of CVEs are now exploited within 24 hours of disclosure. Palo Alto Networks CEO warned organizations have 3-5 months to restructure security posture before AI-driven exploits become normal. (TechPopDaily)

Creator Watchlist

Theo (t3.gg) — "Holy sh*t I think Anthropic is profitable now" (May 28, 28 min) — deep dive on Anthropic's business trajectory (YouTube). Also "Prime is (mostly) right about AI" (May 5) on the AI economy.
No new drops from Chris Raroque, Lennys Podcast, John Kim, or Greg Isenberg in the last 48h window.

Takeaways

The coding agent market is now a two-horse race at scale. Cursor ($3B ARR, potential $60B SpaceX deal) vs. Cognition/Devin ($492M ARR, $26B valuation, 90% internal adoption). Both are growing >100% YoY. If you're building dev tools, you need a thesis on how to coexist with or plug into these platforms.
Regulation is arriving via states, not Washington. Illinois SB 315 passed with near-unanimous support and OpenAI/Anthropic backing. Expect California, New York, and others to follow. Frontier labs are proactively supporting regulation — partly to shape it, partly to create barriers to entry for smaller players.
AI-generated exploits are no longer theoretical. Google confirmed the first real-world case. Combined with 28% of CVEs being exploited within 24 hours, the security landscape has fundamentally changed. AI-native defense tools aren't optional anymore.
Anthropic's path to $1T. $65B raise at $965B valuation with an October IPO on the horizon. Theo's "profitable now" thesis is worth evaluating. The Claude ecosystem (Code, Design, projects) is building a comprehensive platform play.
Sandboxes matter for coding agents. (From queue: Ian's tweet on Modal sandboxes.) As agents write more code autonomously, sandboxed execution environments become critical for safety, reproducibility, and tool isolation. Cognition dogfooding Devin at 90% makes this real, not theoretical.