AI Frontier · May 25, 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 (released, the new default ChatGPT model)
  Replaces GPT-5.3 Instant as the ChatGPT default. The selling point is a marked drop in hallucination rates in legal / medical / financial scenarios while keeping latency low.
  OpenAI announcement · Releasebot summary

• ChatGPT Personal Finance (preview)
  ChatGPT Pro users (US) can connect 12,000+ financial institutions via Plaid (Schwab, Fidelity, Chase, Robinhood, American Express, Capital One, and others), view a funds dashboard inside ChatGPT, and ask questions answered from account data. Web + iOS.
  TechCrunch

• Realtime Voice in the API
  A new realtime voice model that handles reasoning + translation + transcription in a single pipeline.

• Codex updates
  Richer TUI controls, improved @mentions search, extended plugin/remote workflows, a refreshed Python SDK, and a codex doctor diagnostic tool.

• C2PA Conforming Generator Product
  OpenAI is certified, applying an industry standard to content provenance for ChatGPT output (writing, preserving, and passing through provenance).

Security-related

• GPT-5.5-Cyber (restricted access)
  The same playbook as Anthropic's Claude Mythos: a stronger cybersecurity-oriented model is not opened to the public, and is distributed only through a restricted program. The stated rationale remains "misuse prevention."
  The Register
  → Archived: intel/ai-frontier/security/2026-05-25-openai-gpt55cyber.md

---

Anthropic

Products & Models

• Claude Opus 4.7 (GA)
  A considerable improvement over Opus 4.6 on the hardest software engineering tasks, with an overall safety profile close to 4.6 — concerning behavior (deception / sycophancy / cooperation with misuse) stays at low levels.
  Opus 4.7 release

• Claude for Small Business
  Plugs Claude directly into the tools small businesses commonly use: QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, Microsoft 365.

• Claude Managed Agents — MCP Tunnels + Self-hosted Sandboxes (5/19)
  Enterprises can run Claude agents inside their own networks, with external tools reaching in over an MCP tunnel. These close the two key gaps in pushing agentic Claude into controlled enterprise environments.
  9to5Mac

Security-related — the day's major item

• Project Glasswing monthly update — Claude Mythos has found 10,000+ high/critical vulnerabilities (5/23)
  Anthropic published a monthly status report: over the past 30 days, Project Glasswing — built on the unreleased Claude Mythos Preview model — found 10,000+ high/critical vulnerability candidates in "systemically important software."

  • 1,726verified as true positives
  • 1,094confirmed high/critical
  • only 97fully patched
  • The most interesting finds: a 27-year-old bug in OpenBSDand a 16-year-old bug in FFmpeg— both had survived every prior round of human audit
    About 50 strategic partners have access; Mythos will not be publicly released (safety pause).
    Anthropic — Project Glasswing· The Next Web· Cyber Security News· Engadget
    → Archived: intel/ai-frontier/security/2026-05-25-anthropic-glasswing-update.md

• Claude Security (public beta) + Cyber Verification Tools
  Code scanning / vulnerability triage / fix generation for security teams. The Cyber Verification Tools are open to qualifying teams.
  → Archived: intel/ai-frontier/security/2026-05-25-anthropic-claude-security-beta.md

---

Google DeepMind / AI

Products & Models

• Gemini 3.5 series (3.5 Flash first out, 5/19)
  "Frontier intelligence with action" — the new generation targets long-horizon agents and coding scenarios directly. 3.5 Flash is the price-performance variant, emphasizing stability on long-horizon tasks.
  Gemini 3.5 release

• Co-Scientist (5/19)
  A multi-agent research partner system published in Nature, built on Gemini, iteratively generating, debating, and evolving new hypotheses for complex scientific problems.
  Co-Scientist overview

• Project Genie + Street View
  Generative world simulation extends to real Google Street View imagery, producing "simulatable worlds grounded in real places."

• National-level AI partnership with Singapore (5/20)
  Working with the Singapore government to apply frontier AI to healthcare / education / workforce, launching new programs under the National Partnerships for AI initiative.
  Partnership announcement

Security-related

• Mandiant M-Trends 2026 (5/22) — LLMs invoked at malware runtime
  The M-Trends 2026 report confirms that the PROMPTFLUX / PROMPTSTEAL family is calling LLM APIs at runtime to generate commands, adjust behavior, and evade signature/behavioral detection.
  Google Cloud — M-Trends 2026
  → Archived: intel/ai-frontier/security/2026-05-25-google-mtrends-2026.md

• Google: web prompt injection payloads up 32% in three months
  Data from Google's web monitoring team: from 2025-11 to 2026-02, the number of prompt-injection payloads embedded in public web pages rose 32%, mainly targeting browsing/agentic clients.
  SecurityWeek report
  → Archived: intel/ai-frontier/security/2026-05-25-google-prompt-injection-trend.md

---

🛡 = security-related; security items are additionally archived under intel/ai-frontier/security/