AI Frontier · Jun 2, 2026
Product and model developments from OpenAI / Anthropic / Google. Security-related items are marked 🛡 and additionally archived under
intel/ai-frontier/security/.
OpenAI
Products & Models
- No new model releases in the 6/1–6/2 window
The throughline remains preparation of the confidential IPO filing — reportedly with Goldman Sachs / Morgan Stanley, with a confidential S-1 to be submitted within the coming weeks and an IPO as early as September 2026. The Codex desktop agent continues to iterate.
CNBC
Security-related
- ChatGPhish — ChatGPT's web summary rendering weaponized for phishing + indirect prompt injection (ongoing)
Permiso Security disclosed that the chatgpt.com response renderer trusts Markdown links/images from the third-party page it just summarized, automatically fetching attacker-hosted images and rendering links as clickable elements — enabling indirect prompt injection, credential phishing, and leakage of IP / User-Agent / Referer.
The Hacker News
Anthropic
Products & Models
-
Claude Opus 4.8 (5/28) — the current flagship
Stronger than Opus 4.7 on coding, agentic capability, and reasoning; already live in Claude Code with high effort by default.
Releasebot -
Andrej Karpathy joins Anthropic
Per recent reports, Karpathy has joined Anthropic to work on AI-accelerated pre-training research.
LLM Stats -
Code with Claude — Tokyo (6/5–6/6) coming up
The Tokyo leg of the developer conference opens next week.
Anthropic events
Security-related
-
Series H + "Strengthening our model weight security" (5/28)
$65B raised at a $965B post-money valuation; the same day brought updated practices for hardening model weights against exfiltration — against the backdrop of the offensive/defensive AI arms race, "protecting the model weights themselves" has become a frontier security topic.
Anthropic Series H · Model weight security
→ Archived:intel/ai-frontier/security/2026-06-01-anthropic-model-weight-security.md -
Project Glasswing / Claude Mythos — 10,000+ high/critical vulnerabilities (the first monthly update of 5/22 still reverberating)
Mythos has cumulatively found 10,000+ high/critical vulnerabilities in widely used software; the first formal monthly update remains the reference point, and public release stays paused.
Anthropic
→ Archived (already created):intel/ai-frontier/security/2026-05-26-anthropic-glasswing-monthly.md
Google DeepMind / AI
Products & Models
- Gemini 3.5 Flash — the cheaper enterprise-tier push
Google is pushing the low-cost Flash tier hard to win enterprise customers; the Ultra subscription drops to $200/month, with a new $100/month Developer tier. DeepMind absorbed 20+ Contextual AI researchers in a deal worth roughly $80–90M.
HeyGoTrade recap
Security-related
- Google API keys exposed in client-side code can authenticate to Gemini (ongoing)
Google API keys embedded in accessible client-side code (e.g. Maps) can be reused to authenticate to the Gemini AI assistant and access private data. Recommended: rotate keys and tighten client-side key scopes to least privilege.
Bleeping Computer
→ Archived:intel/ai-frontier/security/2026-06-01-google-api-key-gemini.md
🛡 = security-related. Items archived under
intel/ai-frontier/security/are folded into the weekly security roundup each Monday.