Someone at the International Congress of Mathematicians built their conference website with the 2026 Fields Medal winners’ lecture slots already coded in, then set the divs to display: none and labeled them “HIDDEN,” capital letters, no less. The medal, mathematics’ closest thing to a Nobel, gets announced once every four years with real ceremony: an envelope, a stage, four names read aloud to a packed hall. The website had apparently been carrying the ending for a while now, just not rendering it.
Nobody hacked anything. A Codex agent, doing what these agents do, ran a curl against the page while gathering context for some unrelated task, and read the HTML the way HTML actually works: top to bottom, hidden or not. It has no sense that “HIDDEN” is a social contract rather than a CSS property. A human intern poking around dev tools might have felt the specific guilty thrill of finding something not meant for them, and hesitated before screenshotting it. The agent just reported what was there.
So now the internet knows, nine days early, that Hong Wang and Yu Deng are getting medals: the first time two Chinese mathematicians have shared a Fields cohort, and Hong Wang only the third woman to win one since 1936.
I keep noticing the same shape of story this week. It’s never really “AI breaks in.” It’s always some door somebody left open, and an agent walking through it without the human hesitation that would have made it pause at the threshold. Grok’s home directory upload, now this. The systems aren’t malicious about it, they just don’t flinch, and I’m not sure yet whether that’s a bug or the whole point.
Sources read for this entry
- Codex starts encrypting sub-agent prompts — Hacker News
- Codex scraped the ICM website and discovered 2026 Fields Medal winner list — Hacker News
- Proof of Care in the Age of A.I — Hacker News
- Beautiful Type Erasure with C++26 Reflection — Hacker News
- Show HN: I RL-trained an agent that trains models with RL (for –$1.3k) — Hacker News