AI Briefing: May 13 – 15, 2026

Coverage window: May 13 – May 15, 2026 (48 hours)
Generated: 2026-05-15 08:00 UTC
Sources: GitHub API ✅ | Twitter/X API ✅ | arXiv API ✅ | Web Extract ✅ | Web Search ✅ | Wiki Archive ✅

🚨 Breaking (last 24h)

OpenAI Codex Comes to ChatGPT Mobile App

OpenAI launched a preview of Codex inside the ChatGPT mobile app (iOS and Android), letting users manage coding work across laptops, devboxes, and remote environments from their phone. The mobile experience loads live state from the operating machine — active threads, approvals, plugins, and project context — with real-time updates flowing back to the phone (screenshots, terminal output, diffs, test results). OpenAI: Work with Codex from anywhere | TechCrunch | The Verge

Remote SSH is now generally available — Codex connects directly into managed remote environments, auto-detecting hosts from SSH config.
Programmatic access tokens (Enterprise/Business only) — scoped credentials for CI pipelines and release workflows.
Hooks GA — scan prompts for secrets, run validators, log conversations, create memories.
HIPAA-compliant use in local environments for ChatGPT Enterprise workspaces.
Sam Altman announced the launch on X: "Codex in the ChatGPT mobile app!" | Greg Brockman: "Huge step forward for universal usage of agents."

OpenAI Responds to TanStack npm Supply Chain Attack

On May 13, OpenAI disclosed that two employee devices were impacted by the TanStack npm compromise (part of the broader "Mini Shai-Hulud" supply chain attack). Limited credential exfiltration occurred from internal source code repositories accessible to those two employees. No user data, production systems, or intellectual property were compromised. OpenAI: Our response to the TanStack npm supply chain attack

OpenAI is rotating code-signing certificates for iOS, macOS, and Windows as a precaution.
macOS users must update all OpenAI apps by June 12, 2026 — older versions will be blocked by macOS security protections.
The incident occurred during OpenAI's phased deployment of hardened credential controls; the impacted devices lacked the updated configurations.

Anthropic Forms $200M Partnership with Gates Foundation

Anthropic announced a $200 million, four-year partnership with the Gates Foundation focused on global health, life sciences, education, and economic mobility. Anthropic: Gates Foundation partnership

Global health/life sciences (largest component): computational screening for polio, HPV, and preeclampsia therapies; partnership with Institute for Disease Modeling for malaria/TB forecasts.
Education: K-12 tutoring and career guidance in the US, sub-Saharan Africa, and India; public goods (benchmarks, datasets) releasing later in 2026.
Economic mobility: agriculture-specific Claude improvements for smallholder farmers; portable skills records and trustworthy career guidance in the US.

OpenClaw Ships v2026.5.12 Stable + v2026.5.14-beta.1

OpenClaw released v2026.5.12 stable on May 14 (91K+ character release body) and v2026.5.14-beta.1 on May 14 evening. GitHub: v2026.5.12 | GitHub: v2026.5.14-beta.1

Leaner installs: WhatsApp, Slack, Amazon Bedrock, Anthropic Vertex externalized — core installs only pull what you use.
Telegram resilience: isolated polling, durable local spooling, safer group-media handling.
Codex/OpenAI paths smoother: auth-profile-backed media tools, MCP server projection, context-engine thread rotation.
Security hardening: Windows USERPROFILE sandbox blocking, structured env SecretRefs for provider credentials, WiFi ACL audit improvements.
ACP fallbacks: backup runtime backends when primary is unavailable.
Status reactions overhaul: self-explanatory emoji lifecycle (🧠 thinking, 🛠️ tool, ✅ done, ❌ error).
Voice Call/Telnyx: realtime media-streaming call support for conversational voice calls.

📊 Market Moves (last 48h)

OpenAI Codex Free Trial for Companies Switching Over

Sam Altman announced on May 13 that OpenAI is giving companies two months of free Codex usage for the next 30 days to teams switching over. Sam Altman on X | Zeniteq coverage

Greg Brockman noted "great excitement from enterprises wanting to adopt codex."
OpenAI's Codex now has 4 million weekly active users.

Anthropic Claude for Small Business

Anthropic launched Claude for Small Business on May 13 with integrations for QuickBooks, PayPal, HubSpot, and Canva. The SMB Tour started May 14 in Chicago. Anthropic: Claude for Small Business | CIO.com coverage

🔬 Research (last 48h / 7-day arXiv window)

arXiv Papers (May 13, 2026)

15 papers published on cs.AI, cs.LG, cs.CL: arXiv API

WARDEN — Endangered indigenous language transcription/translation (Wardaman → English) with only 6 hours of training data. PDF
EVA-Bench — End-to-end framework for evaluating voice agents across enterprise applications. PDF
Good Agentic Friends — Multi-agent LLM collaboration via weight updates instead of just natural-language messages. PDF
Negation Neglect — Models finetuned on documents flagging claims as false end up believing the claims are true. PDF
History Anchors — Prior behavior steers LLM decisions toward unsafe actions in agent deployments. PDF
Harnessing Agentic Evolution — Iterative candidate generation and feedback for improving programs and workflows. PDF
R-DMesh — Video-guided 3D animation via rectified dynamic mesh flow. PDF
Topology-Preserving Neural Operator Learning — Hodge decomposition for physical field equations on geometric meshes. PDF
QLAM — Quantum long-attention memory for long-sequence token modeling. PDF
Quantifying Sensitivity for Tree Ensembles — Symbolic/compositional approach to verifying decision tree ensemble properties. PDF

🛠️ Tools (last 48h)

Anthropic Python SDK v0.102.0

Released May 13 with cache diagnostics beta and ManagedAgentsSearchResultBlock types. GitHub: v0.102.0

OpenClaw v2026.5.12 Stable Release

The stable release externalizes major provider packages (Bedrock, Slack, Anthropic Vertex), adds ACP fallbacks, overhauls Telegram resilience, and includes a broad security hardening pass. GitHub: v2026.5.12

OpenClaw v2026.5.14-beta.1

Evening release on May 14 adds:

Codex-review skill for maintainers
Voice Call/Telnyx realtime media streaming
Clawdtributor skill for Discrawl-backed contributor PR triage
Text size setting in Control UI
DeepSeek V4 Flash provider docs
Bot loop protection across Discord, Slack, Matrix, Google Chat
Heartbeat scheduler fixes (7 layered issues resolved)

💭 Industry Pulse (last 48h)

Key Twitter Commentary

@sama: "Codex is the best AI coding product and we want to make it easy to try." (May 13) | "Codex in the ChatGPT mobile app!" (May 14) | "also all this:" (May 14 — likely referencing broader product updates)
@gdb: "You can now use Codex, wherever you have it running, from the ChatGPT app. Huge step forward for universal usage of agents." (May 14) | "great excitement from enterprises wanting to adopt codex" (May 13) | "How we built the Codex sandbox for Windows" (May 14)
@AndrewYNg: Launched new course "Transformers in Practice" on May 14 — practical view of transformer-based LLMs for diagnosing problems like slow inference. DeepLearning.AI
@ylecun: Discussed deep net pre-training history with stacked auto-encoders (May 14); "SGD: Stochastic Graduate student Descent" (May 14); "LLMs can actually pulverise asphalt" (May 14)
@karpathy: Last tweet May 11 — "structure your response as HTML" tip.

Hacker News Front Page (May 14)

Claude for Small Business hit the front page.
A Claude Code and Codex Skill for Deliberate Skill Development — community skill for structured learning.

🖼️ New Presentations

No major version update presentations triggered this cycle. OpenClaw v2026.5.12 is a stable maintenance release (not a major version bump); v2026.5.14-beta.1 is a beta. No Hermes Agent releases in the 48h window.

📡 Sources & Data Provenance

Source	Status	URL
Twitter/X API	✅ Working	https://twitterapi.io
GitHub Releases	✅ Working	https://github.com
arXiv API	✅ Working	https://arxiv.org
Wiki Raw Archive	✅ Available	~/wiki/raw/
Web Search	✅ Working	Built-in
Web Extract	✅ Working	Built-in
Hacker News	✅ Working	https://news.ycombinator.com

Each story above links directly to its primary source. Unlinked claims were cross-referenced from multiple sources.