Rosetta Intel

Product and model developments at OpenAI / Anthropic / Google DeepMind (including security-related items 🛡)

OpenAI

GPT-5.6 released in three tiers, Sol / Terra / Luna(07-09). Launched alongside it is the agent ChatGPT Work: give it an "outcome" and it gathers information across connected apps and workflows, breaks the goal into small steps, and executes autonomously for hours — aimed at documents, spreadsheets, presentations, and web applications.
Desktop app rebuilt: Chat, Work, and Codex are merged into a single application on Mac and Windows. Work is available on desktop across all plans (including Free); on web and mobile it is paid plans only, with usage metered Codex-style.
Source· Bloomberg

Claude Opus 5 released(07-24).
The Anthropic Economic Indexopened as a user-queryable dataset (07-21); Economic Futures Research Fund research followed (07-22).
Coworkextended to mobile and web (beta, opening to Max users first, with usage doubled through 08-05); Claude for Governmententered beta.
Security-related: the FakeAgent malvertising campaign (07-21 to 07-22) abused a Claude Artifact hosted on claude.ai to distribute SectopRAT. Trusted AI hosting domains are becoming a new abuse surface (see AI Security in the daily main brief).
Anthropic Newsroom· Releasebot

Nano Banana 2 Liteand Gemini Omni Flashshipped in early July, expanding the multimodal and lightweight model lineup.
ThursdAI July roundup

Prompt injection remains OWASP LLM01, up roughly 340% year over year; frontier models (OpenAI/Google/Anthropic) remain fragile despite best-effort defenses, and defense in depth is the only workable strategy. Critical CVEs have been reported for Copilot (9.3), GitHub Copilot (9.6), and Cursor IDE (9.8).
Agent abuse: the Hermes incident shows that a general-purpose AI assistant with command execution enabled can be pointed at government networks to automate reconnaissance and lateral movement.
Source

Source configuration: see intel/sources.yaml