Rosetta Intel

Product and model updates from OpenAI / Anthropic / Google DeepMind (archived daily). 🛡 = security-related.

OpenAI

GPT-5.5-Cyber released to "trusted defenders" (Daybreak)— OpenAI calls it "our strongest vulnerability discovery and patching model to date," capable of sustained deep analysis over large codebases, validating findings in controlled environments, and developing/testing patches. It set a record 85.6%on CyberGym (which measures whether AI can reproduce known vulnerabilities), versus 81.8% for standard GPT-5.5. Access is limited to verified defenders through Trusted Access for Cyber, with additional monitoring and permission controls.
The Hacker News· OpenAI· Help Net Security
Daybreak Cyber Partner Program— lets 30 cybersecurity vendors/service providers embed GPT-5.5 capabilities into their own products.
BankInfoSecurity
gpt-oss open-weight models— gpt-oss-120b (runs on a single GPU, performance close to o3/o4-mini) and the lighter 20b (runs locally on a high-end laptop).

Claude Fable 5— priced at $10/$50 per million tokens (input/output), with a 1M context window.
Project Glasswing— expands Claude Mythos Preview to roughly 150 organizations (starting from ~50 on 4/7); reportedly it has already helped early partners find 10,000+ high-severity vulnerabilities.
Relayed via FreeBuf weekly

Gemini 3.5 Pro— positioned as June's flagship release (a date has been announced but not finally confirmed); Gemini 3.5 Flashwent live at Google I/O 2026 on 5/19 (API $1.50/$9.00 per million tokens).
Genie 3— a foundation world model that can generate several minutes of interactive 3D environments (720p / 24fps).
Gemini Robotics-ER 1.6— through a Boston Dynamics partnership with Google Cloud/DeepMind, integrated into the Spot robot dog and the Orbit AI visual inspection platform.
LLM Updates· CNBC

Archive: intel/ai-frontier/daily/