Rosetta Intel

OpenAI / Anthropic / Google DeepMind daily updates. 🛡 = security-related.

OpenAI

Partner Network launched(6/14): a new program for ecosystem partners.
New Academy courses(6/12) went live.
Reports say OpenAI plans to acquire Ona; it has also submitted a confidential SEC filing and set up an Economic Research Exchange.
Model status: GPT-5.2 was retired from ChatGPT on 6/12, with existing conversations auto-migrated to GPT-5.5(the current default); frontier models + Codex went live on AWS on 6/2.
OpenAI News· Releasebot

Claude Fable 5 / Mythos 5 released(6/9): adds the new Mythos tier of models.
Claude Opus 4.8(released 5/28): focused on honesty and reliability, more efficient tool calling, and stronger agentic judgment; computer-use hits 84%on the Online-Mind2Web benchmark. The tier lineup is Opus 4.8 / Sonnet 4.6 / Haiku 4.5.
Managed Agentscan now run in customer-controlled sandboxes and connect to private MCP servers (tool execution moves to self-hosted environments).
Enterprise partnerships: TCSand DXCare bringing Claude into regulated industries such as banking and aviation (6/12); 20+ legal MCP connectors and 12 practice-area plugins launched.
Anthropic Newsroom· Releasebot

Gemini 3 family: positioned as a tiered business stack — Gemini 3 Flash (app default), 3.1 Pro, 3.5 Flash, 3.1 Flash-Lite.
Deep Think: a high-intensity reasoning mode for harder tasks.
Deep Research Max: a "step-change" upgrade for autonomous research agents.
Google DeepMind Blog· The Keyword

Anthropic: Managed Agents self-hosted sandboxes + private MCP shrink the exposure surface of agent tool execution (relevant to enterprise security).
Cross-vendor risk: M-Trends 2026 notes that PROMPTFLUX rewrites itself at runtime via the Gemini APIand PROMPTSTEAL (APT28) queries LLMs at runtime — frontier-model APIs are being weaponized as the "brains" of malware. See the AI security / threat intelligence sections of the main brief.

See intel/sources.yaml for configuration