Rosetta Intel

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

OpenAI

GPT-5.6 (Sol / Terra / Luna) went publicly live on 07-09across ChatGPT, the OpenAI API, and Codex, ending a 13-day government-coordinated preview period.
- Sol: flagship tier, adding an Ultra subagent modeand a Maxreasoning-intensity setting.
- Terra: targets GPT-5.5-level quality at roughly half the cost.
- Luna: fast / low-latency tier.
- Greater subagent autonomy means a wider attack surface for agent injection / tool abuse — deployments should pair it with least privilege and human confirmation.
  LLM Stats· ThursdAI 2026-07

Claude Fable 5 returned on 07-01, reclaiming the coding crown among available models with SWE-Bench Pro 80.3%.
- A Mythos-class flagship built for long-horizon agentic operation; priced at $10 / $50 per million tokens (input / output).
- Continues the direction of Claude Sciencefrom earlier this month (aiming to speed up life-science R&D by 10×). Long-horizon agentic operation raises the bar for sandboxing and permission boundaries.
  Fello AI· ThursdAI 2026-07

Gemini 3.5 Pro delayed to 07-17, abandoning the 2.5 Pro architecture for a complete rebuild.
- Introduces a 2-million-token context window, a Deep Think reasoning layer(for complex problem solving), and autonomous workflow capabilities.
- Focused on mathematical reasoning, SVG scene generation, and image quality, aimed squarely at GPT-5.6 and Anthropic Fable 5.
  Bind AI· BigGo Finance

Prompt injection remains the #1 AI security risk of 2026(top of the OWASP LLM list, attack volume +340% year over year). As all three flagships strengthen subagent / long-horizon autonomy, exposure to indirect prompt injection (IPI) and tool abuse expands accordingly; the defensive consensus is "containment": least privilege, human confirmation for risky actions, and tool sandboxing.
CSO Online
The IronWorm payload in the jscrambler npm poisoning specifically harvests configurations from Claude Desktop / Cursor / Windsurf / VS Code / Zed(API keys and MCP server credentials), signaling that the AI development toolchain itself is becoming a high-value target for supply-chain credential theft.
The Hacker News