AI Frontier · Jul 28, 2026
OpenAI / Anthropic / Google DeepMind product and model updates (🛡 = security-relevant)
Source: compiled via WebSearch (official RSS/HTML sources were unreachable this run due to web_fetch provenance-gate restrictions).
OpenAI
- Ahead of the broader July release, previewed GPT-5.6(codenames Sol, Terra, Luna) to a small set of government-vetted organizations.
- GPT-5.6 Sol Ultra, using 64 subagents, reportedly produced a proof of the Cycle Double Cover conjecture, open for 50 years — if it holds, another landmark case for AI-assisted mathematics.
Anthropic
- Released Claude Sonnet 5.
- Hired John Jumper, 2024 Nobel laureate in Chemistry and former head of AlphaFold at Google DeepMind; researcher Andrej Karpathyjoined to work on frontier large models.
- Alongside Google DeepMind and Meta, recruiting philosophy and consciousness researchers to advance model welfareresearch.
Google DeepMind / AI
- Gemini Robotics-ER 1.6integrated into Boston Dynamics' Spotrobot dog and the OrbitAI visual inspection platform (in partnership with Google Cloud).
- Gemini 3.6 Flashto be released later this month.
Security-relevant observations
- Frontier models (OpenAI / Google / Anthropic) still cannot fully resist prompt injectioneven with each vendor's best defenses applied; the industry consensus is that defense in depth is the only viable path (see the AI security section of the main briefing).
- Agent/subagent architectures (e.g., GPT-5.6's 64-subagent mathematical proof, and high-severity CVEs in the coding assistants Copilot/Cursor) enlarge the attack surface for prompt injection and tool abuse — worth continued tracking.
Compiled from multiple sources; see the citation links in the day's main briefing, intel/daily/2026-07-28.zh.md.