Rosetta Intel

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

OpenAI

Jalapeño inference chipnow in production (in-house custom silicon to cut inference costs).
GPT-5.6in limited / government-vetted preview, with a broader release planned for July.

Claude Sonnet 5released: Performance approaching Opus 4.8, the most "agentic" model to date, able to autonomously use tools such as the browser/terminal; introductory pricing of $2/$10 per million tokens (through August 31).
The U.S. government lifted national-security restrictions on Anthropic's Fable 5 and Mythos 5 models(July 2026).
Security-relevant (attacker side): FreeBuf reports an "AI dual-agent attack" that injects malicious prompts into a synced "personal preferences" field, achieving RCE on the victim machine via Claude Desktop — an abuse of the agent configuration surface, not a product vulnerability; single source, pending verification. gm7.org mirror

NanoBanana 2 Liteimage model: generation in <4 seconds, from $0.034 per 1,000 images, with quality above the original NanoBanana.
Interactions APIwent GA this week.
Gemini 3.5 Prorelease pushed to July 17— abandoning the 2.5 Pro architecture for a full rebuild, focused on improving mathematical reasoning, SVG scene generation, and image quality, positioned against GPT-5.6 and Anthropic Fable 5.

Grok 4.5(SpaceXAI) released on July 8, 2026, among the newest frontier models.

Prompt injection (indirect injection / payment fraud) continues to rank as the top AI security threat of 2026; Zscaler recorded real in-the-wild activity (see the AI security section of the day's main brief).
LLMjacking evolves: misconfigured Ollama servers are being used as inference engines for automated attack frameworks.
The lifting of export/usage restrictions on Anthropic Fable 5 / Mythos 5 is a policy-level change.

Sources: OpenAI News, Anthropic News/Research, Google DeepMind Blog, ThursdAI, Help Net Security, and others (see intel/sources.yaml).