Rosetta Intel

Product and model developments at OpenAI / Anthropic / Google, with 🛡 marking security-related items.

OpenAI

gpt-oss open-weight models— the open release includes the 120B(runs on a single GPU, close to o3/o4-mini on several benchmarks) and the lighter 20B(runs locally on a high-end laptop). The open-weight track keeps pressure on the closed-source camp.
Model arms race / IPO— new models are now arriving roughly every 2 days; both OpenAI and Anthropic have filed confidentially to go public, with timing still an open watch item. (Ongoing)

Claude Fable 5 released(2026-06-09) — the first public model in the new "Mythos" family, strong on software engineering, knowledge work, and visual benchmarks, with a 1M-token context window.
Project Glasswing expands— roughly 150 more organizations have been granted Claude Mythos preview access for cyber defense; the program says it has already helped early partners find 10,000+ high-severity vulnerabilities. A representative step in scaling up "AI-assisted vulnerability discovery."

Gemini 3.5 Pro nearing release— a "next month" delivery was promised at I/O; Gemini 3.5 Flash is already GA(the default in the Gemini app and AI Mode in Search, API $1.50 / $9.00 per million tokens). Aimed at agentic/coding use cases.
AI coding rivalry— Microsoft and Google are targeting Anthropic and OpenAI head-on with coding models, making coding one of this round's main battlegrounds.

Project Glasswing (Anthropic)— AI-assisted vulnerability hunting is entering institutional scale, which benefits defenders while also signaling that the same capability can be repurposed by attackers.
Proliferation of open-weight models (gpt-oss)— capable models that run locally lower the barrier to entry for both attackers and defenders; watch the risk of their use for automated exploitation and phishing-content generation.

Configuration: see intel/sources.yaml (ai_frontier category)