AI Frontier · May 31, 2026

Product and model developments from OpenAI / Anthropic / Google. Security-related items are marked 🛡 and additionally archived under intel/ai-frontier/security/.

OpenAI

Products & Models

• GPT-5.5 Instant becomes the new default ChatGPT model (5/5, still widening this week)
  The new default model: on high-stakes medical / legal / financial prompts, hallucinated assertions are down 52.5%versus GPT-5.3 Instant, with inaccurate assertions down 37.3% in especially difficult conversations, plus improved personalization controls.
  TechCrunch· Axios

Security-related

• Rosalind Biodefense — GPT-Rosalind trusted access expands (5/29)
  OpenAI extended trusted access to GPT-Rosalind to vetted developers and US government partners, for biodefense, public health, and pandemic preparedness. The same dual-use, gated-access model approach as GPT-5.5-Cyber — capability released to defenders, backstopped by strong vetting and access controls.
  OpenAI News
  → Archived: intel/ai-frontier/security/2026-05-31-openai-rosalind-biodefense.md

• Content provenance advances — C2PA + cross-platform SynthID watermarking (5/19)
  A layered provenance approach: C2PA conformance, a partnership with Google to apply persistent cross-platform SynthID watermarks to images, and a preview of a tool letting the public verify whether an image came from OpenAI. Infrastructure-layer action against synthetic media misuse.
  Releasebot
  → Archived: intel/ai-frontier/security/2026-05-31-openai-content-provenance.md

---

Anthropic

Products & Models

• Claude Opus 4.8 released (5/28)
  Stronger than Opus 4.7 on coding, agentic capability, reasoning, and practical knowledge work. Claude Code shipped alongside it: high effort by default, dynamic workflows, a faster Fast mode, broader agent / browser / plugin / MCP support, plus a batch of bug fixes and improvements to auto mode.
  Releasebot · Anthropic News

• Series H — $65B raised at a $965B post-money valuation (5/28)
  Led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia, the valuation exceeds OpenAI's for the first time; earlier in May, run-rate revenue crossed roughly $47B.
  Anthropic Series H · Bloomberg

Security-related

• Managed Agents — sandboxes you control + private MCP (same 5/28 window)
  Claude Managed Agents can now run inside enterprise-controlled sandboxes and connect to private MCP servers — both the environment where the agent executes tools and the services it reaches are confined within the enterprise's established boundary. A critical isolation capability for teams running agents in regulated environments.
  Releasebot
  → Archived: intel/ai-frontier/security/2026-05-31-anthropic-managed-agents-sandbox.md

---

Google DeepMind / AI

Products & Models

• Gemini 3.5 Flash + Gemini Omni (continuing from I/O 26)
  3.5 Flash matches frontier agentic / coding work at Flash-class speed, performs strongly on complex long-horizon tasks, and beats Gemini 3.1 Pro on key benchmarks; Gemini Omniis an "any input → any output" multimodal model starting from video generation, billed as a leap in world understanding, multimodality, and editing.
  Google Cloud (I/O 26 innovations)· DeepMind

Security-related

• CodeMender integrated into the Agent Platform
  DeepMind's code security agent CodeMender has been integrated into the Google Agent Platform: autonomously identifying vulnerabilities in code, recommending precise fixes, running security tests, and — once approved — patching across dependent systems. This extends "AI finds holes" into "AI fixes holes and lands patches across dependencies."
  DeepMind
  → Archived: intel/ai-frontier/security/2026-05-31-google-codemender-agent-platform.md

• Big Sleep — sustained push on the defensive side
  Big Sleep (the AI vulnerability discovery agent from DeepMind / Project Zero) remains the first AI agent on public record credited with directly thwarting an in-the-wild exploitation attempt (see the 5/11 AI-developed 2FA bypass 0-day intended for mass exploitation). The arms race between offensive and defensive AI tooling is playing out in real time.
  Google Cloud blog
  → Archived: intel/ai-frontier/security/2026-05-31-google-bigsleep-defense.md

---

🛡 = security-related. Today's security items are also archived under intel/ai-frontier/security/.