Rosetta Intel

Product and model updates from OpenAI / Anthropic / Google. 🛡 = security-relevant.

OpenAI

GPT-5.6 family generally available— The Sol / Terra / Luna variants went GA on Jul 9 (previewed under a U.S. government access list on Jun 26). OpenAI is reportedly preparing to release a new open-weight language model.
LLM updates· SeekingAlpha

Claude Opus 5 released (Jul 24)— Topped the Artificial Analysis intelligence index immediately upon release; Claude Sonnet 5 had already become the default model on Jun 30.
ThursdAI July 2026
Jailbreak-resistance comparison— Axis Intelligence testing as of July 2026 shows Claude Opus 4.7 with the lowest aggregate vulnerability count and the highest single-turn jailbreak resistance among the frontier models tested; however, all frontier models (including OpenAI, Google, and Anthropic) retain residual prompt-injection risk even with best defenses applied.
AI Model Vulnerability Tracker

Gemini 3.6 Flash released (Jul 21)— Cheaper and faster, now powering the free Gemini app; launched alongside Nano Banana 2 Lite (fastest, most cost-efficient image generation) and Gemini Omni Flash (video + conversational editing).
ThursdAI July 2026
Gemini 3.5 Pro still delayed— Reportedly postponed due to an architecture rebuild; not yet released, months behind schedule.
BigGo Finance

Frontier model capabilities keep iterating rapidly (27 launches this month), but prompt injection remains an unsolved systemic risk in agentic / RAG / multimodal / coding-assistant scenarios (OWASP LLM01, 50–84% success rates).
With the EU AI Act's August 2026 deadline approaching, compliance and red-teaming pressure on model providers is rising.
Prompt injection overview (Vectra)

Configuration: see the ai_frontier / ai_security sections of intel/sources.yaml