Rosetta Intel

OpenAI / Anthropic / Google product and model news; 🛡 marks security-relevant items.

OpenAI

GPT-5.6 previewed(June 26) as a three-tier family: Sol(flagship), Terra(balanced mid), Luna(fast/cheap). Access still limited to ~20 government-vetted partner orgs this week; broad availability expected mid-to-late July.
Cerebras inference plan— OpenAI intends to run Sol on Cerebras wafer-scale hardware at up to 750 tokens/sec, ~15× current GPU-based inference throughput.
Governance— OpenAI's proposed 5% government stake is framed as part of voluntary engagement with a frontier-model standards framework ahead of a planned September 2026 IPO.
llm-stats· AI Tools Recap

Claude Enterprise admin controls(July 3) — richer analytics, model-level entitlements, and spend alerts for enterprise admins.
Claude Fable 5 fully restored globally(from July 1) after the U.S. Commerce Department lifted the 19-day export-control suspension (triggered by an Amazon-reported jailbreak showing vuln-finding / exploit-generation). Restoration ships a new safety classifier that blocks the flagged technique in >99%of cases.
AI Tools Recap· The Hacker News

Gemini 3.5 Pro slipped to July— promised at I/O in May for a June release, delayed after enterprise testers flagged reasoning/coding regressions.
Gemini 3.5 Flash shipped on timeand outperforms the previous generation.
llm-stats

First agentic ransomware (JADEPUFFER)— Sysdig documents an LLM-driven end-to-end database-extortion operation (Langflow RCE CVE-2025-3248 → Nacos/MySQL), underscoring that frontier-model capability now directly lowers the ransomware skill floor. Full write-up in today's main brief (AI Security).
Sysdig
White House AI release standards— advanced talks with OpenAI, Google, and Anthropic on voluntary release standards; an announcement is reportedly possible in the coming week.
AI Tools Recap