AI Frontier · May 20, 2026
Weekly roundup (May 13 – May 20) of developments from OpenAI, Anthropic, and Google
🛡 marker = security-related, also archived to../security/
OpenAI
| Date | Topic | Notes |
|---|---|---|
| 5/5 | GPT-5.5 Instant launched, replacing 5.3 Instant as the ChatGPT default model | Hallucination rate in medical/legal/financial scenarios 52.5% lower than 5.3 Instant; available via API as chat-latest |
| 5/7 | New Voice API model | Can reason / translate / transcribe speech |
| 5/14 🛡 | Codex arrives in the ChatGPT mobile app | Remote SSH, hooks, access tokens, and HIPAA support launched for the enterprise tier |
| 5/15 | Personal Finance feature | Available to Pro users in the US; connects bank accounts |
Links:
Worth digging into:
- With SSH + hooks, Codex Mobile means ChatGPT can now execute code remotely from a phone — how should this be assessed from an agent-security perspective?
- The Finance feature connects bank accounts — how does OpenAI handle OAuth / credential security?
Anthropic
| Date | Topic | Notes |
|---|---|---|
| 5/7 🛡 | The Anthropic Institute (TAI) launched | Four research directions: economic diffusion / threats and resilience / AI systems in the wild / AI-driven R&D |
| Ongoing 🛡 | Project Glasswing + Claude Mythos Preview | $100M in model credits; Mythos autonomously discovered "thousands of high-severity vulnerabilities" covering all major OSes and browsers; partners include AWS / Apple / Microsoft / Google / CrowdStrike / Palo Alto + ~40 organizations |
| Ongoing 🛡 | Anthropic Fellows Program applications open for the May + July 2026 cohorts | Covers scalable oversight / adversarial robustness and AI control / model organisms / interpretability / AI security / model welfare |
| Roadmap 🛡 | Frontier Safety Roadmap | Companion to the RSP |
Links:
- Project Glasswing
- Anthropic Institute Agenda
- Frontier Safety Roadmap
- Platformer analysis
- Fortune industry reaction
Worth digging into:
- What is the actual distribution of the "thousands of 0-days" Mythos found? Which categories? Will any portion be published for research?
- How do researchers outside Glasswing's 41 partners get access? Will this create a new asymmetry (those with access can fix, those without become targets)?
Google DeepMind / AI
| Date | Topic | Notes |
|---|---|---|
| 5/19 (I/O) | Gemini Omni — multimodal video generation / editing | Omni Flash live for AI Plus/Pro/Ultra subscribers and free YouTube Shorts/Create users |
| 5/19 (I/O) | Antigravity + Gemini 3.5 Flash | Agentic coding & long-horizon tasks; demoed building a working OS from scratch in 12 hours |
| 5/19 (I/O) | Ultra subscription $250 → $200/mo; new Developer tier at $100/mo | |
| 5/19 (I/O) | DeepMind acquires 20+ Contextual AI researchers ($80–90M licensing deal) | |
| May | AI-Enabled Pointer research project surfaces | Understands not just the "content" a user points at, but the "intent" |
| Ongoing 🛡 | Big Sleep (Project Zero + DeepMind) | AI agent autonomously discovering real vulnerabilities; has stopped one flaw that was about to be exploited by a threat group |
Links:
Worth digging into:
- Antigravity + Gemini 3.5 Flash can build an OS in 12 hours — how should the agent supply-chain attack surface be assessed?
- Gemini Omni's "riskiest feature" for video generation was held back — what is it?
Cross-vendor observations
- All three are pushing agentic / long-horizon capabilities: Anthropic uses Mythos to find vulnerabilities, OpenAI brings Codex to mobile + enterprise SSH, Google uses Antigravity to build an OS. The agent attack surface is expanding at unprecedented speed.
- The OWASP Top 10 for Agentic Applications (published 2025/12)arriving at this moment makes perfect sense — the gap between taxonomy and practice keeps widening.
- AI finding vulnerabilities vs AI exploiting vulnerabilities: Big Sleep and Mythos on the defensive side; PROMPTFLUX/PROMPTSTEAL (Mandiant report) on the offensive side. The same underlying technology.