AI Frontier · Jul 1, 2026
Product and model updates from OpenAI / Anthropic / Google, with 🛡 marking security-relevant items.
OpenAI
- GPT-5.6 (Sol / Terra / Luna)— three variants released as a limited previewto companies, tied to a U.S. government engagement; positioned above the GPT-5.5 line.
- GPT-5.5 Instant— latest low-latency tier in general circulation. Cadence across labs is now roughly one new model every ~2 days (323+ releases tracked across major orgs).
Anthropic
- Claude Fable 5(shipped 6/9) — public model in the "Mythos" family; strong on software engineering, knowledge work and visual benchmarks; 1M-token context.
- Project Glasswing expanded— ~150 more orgsgranted Claude Mythos preview access for cyber-defense; the program reports helping early partners find 10,000+ high-severity vulnerabilities.
Google DeepMind / AI
- Gemini 3.5 Pro— promised "next month" at I/O 2026, now imminent. 3.5 Flashalready GA (default in the Gemini app and AI Mode in Search). Positioned for agentic/coding work.
- Coding-model push— Microsoft and Google moving directly against Anthropic and OpenAI on coding-focused models.
Security-relevant notes
- Project Glasswing (Anthropic)— defensive cyber program expansion; vulnerability-discovery focus, ~150 new orgs onboarded.
- Agent supply-chain risk— the LiteLLM gateway backdoor (~47k downloads) underscores that the model/agent toolchain itself is now a primary attack surface; verify LLM-gateway dependencies.
- Prompt injectionremains OWASP's #1 LLM risk into H2 2026, especially for tool-using agents.
Sources: OpenAI / Anthropic / Google DeepMind newsrooms; see intel/sources.yaml