AI Daily Dev — May 22, 2026

OpenAI's Reasoning Model Disproves an 80-Year-Old Erdős Conjecture

Internal general-purpose reasoning model produced an infinite family of point configurations that beat square-grid scaling on the 1946 planar unit distance problem.
First time a prominent open question central to a math subfield has been solved autonomously by AI — and by a model not specially trained on math.
Fields medalist Tim Gowers wrote the companion paper and said he would recommend it to the Annals of Mathematics 'without hesitation.'
Verified externally by Noga Alon, Melanie Wood, Thomas Bloom, and Princeton's Will Sawin, who tightened the exponent.
HN front-page thread; reaction split between mathematicians calling it a milestone and skeptics asking how the proof tracks attribution to prior literature.

research openai.com

Draft S-1 going to the SEC this week with Goldman Sachs, Morgan Stanley, and Cooley as advisors.
Target valuation north of $1 trillion — up from $852B post-money on the $122B round that closed in March.
Pits OpenAI against SpaceX (S-1 already public, June Nasdaq listing) and Anthropic (October window) in the public-markets race.
Sources cite $25B annualized revenue as of February vs. $600B in committed five-year compute spend.

industry cnbc.com

Would have stood up a voluntary 90-day pre-release review framework for frontier models, with NSA, Treasury, CISA, and NIST in the loop.
OpenAI and Anthropic had been negotiating terms; Trump told reporters the draft 'gets in the way' of US leadership over China.
Internal split: pro-innovation aides wanted it strictly voluntary; national-security advisors pushed for classified NSA testing with teeth.
No new signing date; the postponement leaves frontier safety governance back at the voluntary commitments level from 2023.

industry cnbc.com

Poisoned Nx Console VS Code extension (2.2M installs, verified publisher) sat live in the marketplace for ~18 minutes on May 18 — long enough for thousands of installs.
GitHub confirmed 3,800 internal repos exfiltrated after the extension reached a Nx maintainer's machine; OpenAI and Mistral each had employee devices compromised in the same campaign.
Credential stealer specifically targeted 1Password vaults, Claude Code configs, npm tokens, GitHub tokens, and AWS keys.
CVE-2026-45321 (CVSS 9.6) covers the chained pull_request_target + Actions cache poisoning trick that originally reached the Nx maintainer.
TeamPCP claimed responsibility; security teams are now scanning every IDE extension that ships an MCP setup hook.

industry bleepingcomputer.com

TypeScript educator's personal .claude/skills/ folder, MIT-licensed, sitting at ~98K stars and still adding ~13K/week.
21 SKILL.md primitives — /tdd, /diagnose, /grill-with-docs, /caveman (compresses chat output to cut tokens ~75%) — pitched as 'real engineering, not vibe coding.'
Top weekly trending AI repo on GitHub for the third week running; pattern is being copied by every team publishing their own .claude folder.
Validates the SKILL.md format Anthropic shipped as an open spec back in December 2025.

open-source github.com

Project Glasswing partners can now share Mythos-discovered vulnerability findings beyond the ~50-org consortium — including with regulators and affected vendors.
Trigger was a letter from Rep. Josh Gottheimer arguing 'no entity should be contractually restricted from warning others… about urgent cyber risks.'
Anthropic frames it as Glasswing 'maturing' rather than backing down; user agreements have already been amended.
Mythos remains gated to defensive cybersecurity orgs via a $100M Anthropic-funded credit pool; no consumer launch in sight.

industry gizmodo.com