JadidJust-in AI Developments, Incidents & Deadlines
May 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

AI news, Saturday, May 23, 2026

11 stories

TLDR

  • OpenAI internal reasoning modelA model that works through hidden reasoning steps before its final answer. Those steps are billed as output.OpenAI ↗ autonomously disproves an 80-year-old Erdős discrete-geometry conjecture (May 20) — Fields medalist Tim Gowers calls it "a milestone in AI mathematics." First time a generalist model (not a math-specialized system) has settled a central open problem in a subfield of math, not just a competition benchmark.
  • Anthropic's Project Glasswing has surfaced 10,000+ high/critical CVEsCommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ across widely-used software in one month (May 22) — Cloudflare ~2,000 bugs (400 critical), Mozilla 271 fixes in Firefox 150; Anthropic says the bottleneck is now disclosure and patching, not finding.
  • Andrej Karpathy joins Anthropic's pretraining team (May 19) — OpenAI co-founder and former Tesla Autopilot lead will use Claude to accelerate pretraining research under Nick Joseph. Strongest signal yet that Anthropic is betting on AI-assisted research velocity vs. raw compute.
  • Bristol Myers Squibb deploys Claude Enterprise to all 30,000 employees (May 20) — Claude becomes BMS's "shared intelligence platform" across research, clinical, manufacturing, commercial, and engineering; Claude Code is the explicit engineering surface.
  • Leaked Zuckerberg all-hands audio: Meta's "Model Capability Initiative" surveilled employees to train AI (Audio surfaced May 21–22, meeting April 30) — Mouse, keystrokes, and screenshots from Gmail, GChat, VSCode and Metamate; 1,000+ internal signatures on a petition to halt it after this week's 8,000-person layoff.

Action Items

Things your engineering team should look at or do this week.

  • Bake Glasswing-grade dependency hygiene into next sprint — Anthropic's May 22 update reports >10,000 critical CVEs found via Mythos Preview across systemically important OSS, with partners reporting 10× bug-discovery gains and Anthropic explicitly working toward a Mythos-class general release. Two same-day moves: (1) inventory the OSS in your bill-of-materials most likely to be in scope (browsers, kernels, popular libs) and patch the latest CVE waves now, before Mythos goes GA; (2) audit your CVE intake + disclosure pipeline — Anthropic flagged that verification and disclosure, not detection, is the new bottleneck, so an inbox-of-PRs problem is coming. (Anthropic blog | Engadget)

  • Re-run your Codex vs. Claude Code bake-off — Goal Mode is now GA — OpenAI shipped Goal Mode to GA across the Codex CLI, IDE extension, and Codex App on May 22; macOS users also got "Appshots" (hotkey attaches an app window + visible text into a Codex thread). Both materially change the developer loop. If you've been holding off picking a primary coding agent for the org, this is the moment to re-evaluate. (OpenAI release notes)

  • Write down your internal "AI surveillance and training-data" policy this week — The Meta leak and the back-to-back Glasswing/Mythos coverage will land in engineers' inboxes. Two artifacts to produce now: (a) what monitoring/telemetry leaves your systems to outside AI vendors (Copilot, Claude, Cursor, ChatGPT, etc.); (b) what activity data your own internal AI may or may not consume. Treat this as on par with your data-retention policy. (eWeek on Meta)

Models & Tools

  • OpenAI Codex — Goal Mode GA, macOS Appshots, browser annotations (May 22) [Agents] — Several long-running-task and context features in one drop. Goal Mode (define an outcome + success criteria; Codex iterates until met) is GA across CLI, IDE extension, and Codex App. macOS Appshots: a hotkey attaches a chosen app window with screenshot + visible text into a Codex thread, eliminating long setup prompts. In-app browser annotations also got more precise for frontend work. Separately, an elevated-latency/error window on ChatGPT 5.5 Thinking resolved earlier the same day. (OpenAI release notes)

  • Anthropic Claude Code — terminal/Windows/PowerShell fix sweep + new managed-agent features (May 22) [Agents] — A maintenance release with smarter pinned background sessions (stay alive when idle; restart in place to apply updates), a renamed /code-review flow, and a fix for the Bash tool exit-code-127 regression introduced in 2.1.147. Alongside the Code drop, Anthropic also published higher Claude Code + Opus API limits and three new Claude Managed Agents capabilities: multi-agent orchestration (coordinator spins up parallel subagents), "dreaming" (agents learn from past sessions between runs), and outcomes (run an agent in a loop until a stated outcome is met). (Anthropic — higher limits + agents | Claude Code GitHub releases)

  • WebMCP — Chrome 149 origin trial opens to developers (May 18–19) [Agents] — Google's I/O proposal for client-side MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ — websites expose JavaScript functions and HTML form tools in-page so in-browser agents (today, Gemini in Chrome) can call them without a backend MCP server — graduated from behind-a-flag to a public origin trial. Docs landed May 18, the keynote announcement came May 19. Spec lives in the W3C Web ML Community Group, not yet on the Standards Track. Sign up if you'd rather expose structured tools than have agents scrape your DOM. (Chrome for Developers | Chrome at I/O 2026)

  • Salesforce Agentforce Coworker — beta GA across all Agentforce editions (May 21) [Customer AI] [Agents] — Marc Benioff's Agentforce Coworker puts an AI teammate "in every search bar" — it taps CRM data, workflows, opportunity and case history, and acts in real time, surfaced inside Salesforce, Slack, Teams, ChatGPT, and beyond. Available immediately for Enterprise, Unlimited, and Agentforce 1 customers; flagged as the fastest single-product setup users have seen on the platform. Notable as Salesforce's most aggressive packaging yet of agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ features into the core sales motion. (Salesforce Ben | Salesforce Developers docs)

Security

  • Project Glasswing — 10,000+ critical vulnerabilities found, partner pool expanding (May 22) [Trending] — Anthropic's first formal update on Glasswing since the April launch. Mythos Preview + ~50 partners surfaced >10,000 high/critical-severity vulnerabilities in widely-used software in roughly a month. Cloudflare: ~2,000 bugs (400 critical) with fewer false positives than human-led testing. Mozilla: 271 vulnerabilities fixed in Firefox 150; Mythos noted as materially more effective than Claude Opus 4.6 on the same workload. Anthropic explicitly says verification and disclosure is now the bottleneck and is engaging US + allied governments to scale the program. Expect an unusual wave of CVE disclosures across popular OSS over the next 6–12 months, even before Mythos-class models go GA. (Anthropic blog | Engadget | Cybersecurity News)

Industry & Business

  • OpenAI's internal reasoning model disproves Erdős's unit-distance conjecture (May 20) [Trending] — OpenAI published a research note: an internal general-purpose reasoning model (not a math-specialized system) found an infinite family of constructions that beats the long-standing square-grid bound on Erdős's 1946 unit-distance problem, using deep algebraic number theory (Golod–Shafarevich, infinite class-field towers). Fields medalist Tim Gowers called the result "a milestone in AI mathematics." First case of a generalist AI autonomously settling a central open problem in a subfield of math, distinct from benchmark or olympiad-style tasks. (OpenAI announcement | Scientific American | Interesting Engineering)

  • Andrej Karpathy joins Anthropic's pretraining team (May 19) [Hiring] — OpenAI co-founder, former Tesla Autopilot lead, and Eureka Labs founder Karpathy is starting a new team that uses Claude to accelerate pretraining research, under Nick Joseph. Pretraining is Anthropic's most compute-intensive phase, and the framing — AI accelerating its own training — is the clearest signal yet that Anthropic believes research velocity, not raw flops, is its lever vs. OpenAI and Google. (TechCrunch | CNBC | Axios)

  • Bristol Myers Squibb — Claude Enterprise across all 30,000 employees (May 20) [Customer AI] — Multi-year strategic agreement positions Claude as BMS's "shared intelligence platform" across research, clinical development, manufacturing, commercial, and corporate functions, with Claude Code as the explicit engineering surface. Named use cases: drafting clinical study reports from underlying trial data, regulatory submission support, retrieval over decades of proprietary scientific data, real-time root-cause investigation of manufacturing deviations. One of the largest enterprise AI commitments to date in pharma. (BMS press release | Quartz | MobiHealthNews)

  • Leaked Zuckerberg all-hands audio reveals Meta's "Model Capability Initiative" (Audio surfaced May 21–22; meeting was April 30) [Trending] [Hiring] — Audio obtained by More Perfect Union captures Zuckerberg defending an internal program that collects mouse clicks, keystrokes, and periodic screenshots from approved apps (Gmail, GChat, VSCode, and Meta's internal Metamate) to train coding models. Stated rationale: Meta engineers are "smarter than the contractors the rest of the industry relies on." The leak landed against this week's 8,000-person layoff; 1,000+ Meta employees have signed an internal petition to halt the program. (eWeek | Common Dreams)

AI in Production

  • Claude Opus 4.7 & Haiku 4.5 + OpenAI Codex — same-day partial outages (May 22 morning) [Production] — Anthropic's status page logged elevated errors on Opus 4.7 and Haiku 4.5 between roughly 10:26–10:31 UTC, fully resolved by mid-morning. OpenAI reported a partial outage across Codex App, Codex API, and Codex CLI in the same window, plus elevated latency on ChatGPT 5.5 Thinking. Both fully resolved by end of day. Worth a quick check if synthetic monitors fired against either platform yesterday. (Claude status | OpenAI status)

Policy & Safety

  • Jack Clark (Anthropic co-founder) at Oxford: AI-assisted Nobel within 12 months — and "non-zero chance it kills everyone" (May 21) [Trending] — At Oxford's Institute for Ethics in AI, Clark predicted: AI-assisted Nobel-worthy discovery within a year; bipedal trade robots within two; AI-run companies generating millions in revenue within 18 months; recursive self-improvement by 2028. He paired the optimism with a deliberate restatement of catastrophic risk: there remain "plausible scenarios" in which the technology has a "non-zero chance of killing everyone on the planet," and "that risk hasn't gone away." TIME's read: Anthropic explicitly selling capability while reasserting safety. (TIME | TechCentral.ie)

Type to search · ↑↓ to move · enter to open