How the Copilot Worm and the Munich Suno Ruling Rewrote AI's Legal and Security Calculus

The Day's Thesis

▶

Signal of the Day: Amazon raised FY2026 capex guidance to ~$220 billion — up from ~$200 billion — as AWS customers are already contracting capacity through 2028 on five-year AI agreements, even as a self-spreading Copilot worm remained unpatched for 144 days and a Munich court ruled Suno liable for copyright infringement in both training and output.

The 30-Second Read:

• Amazon's FY2026 capex guidance hits ~$220B, with AWS reporting 37% cloud growth and a 39.4% operating margin
• Microsoft Copilot worm — unpatched after 144 days and two attempts — demonstrates systemic prompt-injection vulnerability in production enterprise AI
• Munich court rejects fair use and text-and-data-mining defenses for Suno, finding six copyrighted songs reproducibly stored in model weights
• Meta's free cash flow collapsed 91% to $784M against $30B in Q2 capex, with full-year guidance raised to $130–$145B

The day's news reveals a structural tension: hyperscaler infrastructure commitments are accelerating past $200B annually even as legal and security liabilities in deployed AI systems remain unresolved — a combination that will force enterprise procurement and legal teams to price risks they have largely ignored.

---

AI & Research Frontier

GPT-5.6 Pro solved two open mathematics problems on first attempt, each previously worked on at length by Fields Medal winner Timothy Gowers — a capability threshold that has no clear precedent in the discipline.

OpenAI's systems have now refuted the Unit Distance Conjecture and addressed two Gowers problems, compressing research cycles that previously required months of expert effort.

Gowers publicly warned of the "possible destruction of mathematical culture" if practitioners cede the expertise needed to verify such results — a concern with direct operational weight, since a parallel OpenAI field report found coding agents capable of 60x software speedups but "eloquent, convincing, and confidently wrong in ways that are easy to miss." The verification burden shifts entirely to human reviewers, a labor cost that scales poorly.