AI Frontier · Jun 4, 2026
Daily developments from OpenAI / Anthropic / Google DeepMind · Window: past 24 hours (2026-06-03 → 2026-06-04)
🛡 = security-related
OpenAI
- Role-specific Codex plugins + Sites + annotation (6/3)— OpenAI released 6 no-code plugins aimed at knowledge work, covering 62 common apps and 110 skillsin total: the data analysis plugin connects to Snowflake, Databricks Genie, Hex, and Tableau to explain metric changes and generate reports/dashboards; Sites(sharing interactive workspace apps and websites) is in preview; and new in-place annotationlets teams refine documents, spreadsheets, and slides directly. Codex weekly actives have reached 5 million. SD Times roundup· Releasebot
- ChatGPT "Active sessions" security feature— users can review account sessions and sign out ones they don't recognize, with device, location, sign-in time, trusted status, and current-session details shown. A practical reinforcement on the account security side. Releasebot
- Model retirement schedule (reminder)— GPT-4.5 exits ChatGPT on 6/27(30-day sunset); o3 exits on 8/26(90-day sunset). OpenAI model release notes
Anthropic
- Services Track + Partner Hub launched (6/3)— Anthropic announced a new Services Trackin the Claude Partner Network (a tier for services/integration partners) plus a unified Partner Hub, alongside partner ecosystem expansion. Anthropic news
- Project Glasswing expands to about 150 new organizations across 15+ countries (6/2)— bringing total partners to roughly 200; newly added are previously thin critical infrastructure sectors including power, water, healthcare, telecommunications, and hardware. Anthropic estimates that for most Glasswing partners, a major attack on their codebase could affect 100 million+ people. It also launched Claude Security(codebase scanning + patch recommendations powered by publicly available frontier models such as Opus 4.8), and opens Glasswing's internal tooling to trusted security teams on request. Anthropic — Expanding Glasswing· TechCrunch· Cybersecurity News
- Claude Code ships Opus 4.8— high effort by default on hard tasks; introduces dynamic workflows(letting Claude create workflows and orchestrate dozens to hundreds of agents in the background); a faster Fast mode. Releasebot
- A note on dual-use capability— the same week, Sophos disclosed an AI-built ransomware toolkit whose development leaned on Claude Opus + Cursor agents(including automated AD reconnaissance and EDR bypass labs). As frontier coding capability rises, offensive misuse rises with it — see the AI security section of the main brief. Help Net Security
- IPO context (reminder)— Anthropic filed confidentially for an IPO between 5/28 and 6/1, having previously closed a $65B Series H at a roughly $965B valuation, with annualized revenue reportedly around $47B. CBS News
Google DeepMind / AI
- Gemini 3.5 Flash GA + Gemini Omni + Deep Think rollout— Gemini 3.5 Flashis positioned as the strongest agentic/coding model currently available, beating Gemini 3.1 Pro on several key benchmarks and performing strongly on long-horizon tasks; Gemini Omnidelivers any input to any output (starting from video); Gemini 3 Deep Thinkis available to Google AI Ultra subscribers as the top-tier reasoning mode. The Ultra subscription drops to $200/month, with a new $100/month Developer tier. Google Cloud — I/O 26· llm-stats
- DeepMind recruits 20+ Contextual AI researchers— reportedly an $80–90Mlicensing/talent deal, continuing the industry-wide fight over top AI talent. HeyGoTrade
- Continuing watch: the GTIG/DeepMind report on "Gemini used across all attack phases"— yesterday's detailed coverage of Honestcue runtime AI malware and APT42 phishing personas is still developing, and together with this issue's Sophos AI ransomware toolkit forms a coherent narrative of a steadily expanding AI offensive surface. The Record
This file is the standalone AI Frontier archive; the main brief is at intel/daily/2026-06-04.{zh,en}.md