Rosetta Intel
Briefings/Daily BriefAI FrontierRansomware
Rosetta Lab ↗Blur Horizon LLC
AI Frontier2026-05-23
AI Frontier·2026-05-23

AI Frontier · May 23, 2026

Same-day developments from the three vendors: OpenAI / Anthropic / Google DeepMind.
Source: WebSearch synthesis (feed egress restricted; see the "failed sources" section of the main brief).

Today's highlights

VendorUpdates today?Focus
OpenAI✅ OngoingGPT-5.5 Instant still rolling out in batches; C2PA Conforming Generator coverage continues
Anthropic✅ OngoingKPMG / PwC alliance expansion in progress; Code with Claude 2026 content still being released
Google DeepMind✅ Post-I/O momentumGemini 3.5 Flash rollout; Gemini Spark beta list expanding; DeepMind hiring wave

OpenAI

Overall status: A dense May product cadence (GPT-5.5 Instant / Voice Intelligence API / Codex for business / personal finance / C2PA); no new "big news" as of 5/23, but earlier items keep penetrating the enterprise / consumer sides.

  • GPT-5.5 Instant (released 5/5, rollout continuing)
    ChatGPT default model; markedly lower hallucinations in legal / medical / financial domains; improved personalization controls. Low latency retained.
    Releasebot — OpenAI · TechCrunch

  • Voice Intelligence API (5/7)
    Real-time voice model: reasoning + translation + transcription chained within a single API.
    Releasebot — OpenAI

  • Codex business workflows (5/14–5/15)
    On the ChatGPT Business side: appshots, goal mode goes GA, browser annotation + browser-use tools, remote locked computer use, shared workspace plugins, admin adoption analytics.
    Releasebot — OpenAI

  • ChatGPT Personal Finance (5/15 preview)
    Available to US Pro users; Plaid integration reaching 12,000+ financial institutions (Schwab, Fidelity, Chase, Robinhood, AmEx, Capital One, etc.); Q&A spans spending analysis → future financial planning.
    TechCrunch

  • 🛡 C2PA Conforming Generator Product (5/19)
    OpenAI formally passes C2PA conformance certification; provenance metadata attached to OpenAI-generated content can be read, preserved, and propagated by platforms. One of the machine-readable foundations for countering AI deepfake disinformation.
    Releasebot — OpenAI


Anthropic

Overall status: Dense enterprise-side moves in May (KPMG global alliance, PwC expansion) while the developer platform keeps shipping new capabilities (Managed Agents, self-hosted sandboxes); no new model release as of 5/23, but earlier items are still rolling out.

  • Claude for Small Business (5/13)
    Bundle: Intuit QuickBooks / PayPal / HubSpot / Canva / Docusign / Google Workspace / Microsoft 365 connectors + ready-to-use workflows.
    Anthropic news · SiliconANGLE

  • KPMG global alliance
    Claude embedded in KPMG's "Digital Gateway"; new tools for tax / legal clients; reaching 276,000+ employees worldwide.
    Anthropic announcement · KPMG press

  • PwC partnership expansion
    Claude Code + Cowork rolled out to PwC US teams; heading toward hundreds of thousands of professionals globally.
    Anthropic announcement

  • Code with Claude 2026 (5/6, San Francisco; content continuing to publish)
    Launched Managed Agents capabilities: proactive workflows, dreaming, multi-agent orchestration, outcomes, webhooks; Claude Code + Opus API limits raised; self-hosted sandboxes in public beta, keeping sensitive files / packages / services on users' own infrastructure.
    InfoQ

  • 🛡 Project Glasswing + Claude Mythos Preview (ongoing from 5/19–5/22)
    Mythos Preview continues find/fix work on the partner side; details of the 500 zero-days are in yesterday's (5/22) main brief AI Security section. No new public material added today.


Google DeepMind / Google AI

Overall status: Google I/O 2026 afterglow continues (5/19 main keynote, 5/20 follow-up sessions); Gemini 3.5 Flash / the Ultra price cut / the new Developer tier are all in rollout; the Gemini Spark beta is expanding.

  • Gemini 3.5 Flash released; Ultra subscription cut from $250/mo to $200/mo; new $100/mo Developer tier (for engineers / professional users).
    Heygotrade I/O recap · eWeek

  • Gemini Spark
    General-purpose AI agent that reasons over information across connected apps; beta for trusted testers + Google AI Ultra subscribers.
    CNBC

  • Omni world model
    Physical-environment simulation; entering Google's AI product matrix.
    CNBC

  • Gemini Intelligence on premium Android + Chrome (Android Show: I/O Edition, 5/12) + Googlebook, a Gemini-first laptop category teased.
    Sources.news preview

  • User numbers
    Gemini at 900M MAU (2× YoY; 400M last May); AI Overviews on Search at 2.5B MAU.

  • DeepMind hiring
    20+ Contextual AI researchers hired via an $80–90M licensing deal. The agent / context research talent pool keeps being absorbed.

  • 🛡 Project Zero — "A 0-click exploit chain for the Pixel 10" (5/13)
    Two vulnerabilities in the /dev/vpu driver (Chips&Media Wave677DV / Tensor G5) on the Pixel 10 chained into a zero-interaction escalation to root.
    Project Zero original · CybersecurityNews


Security-related items (also archived separately)

  • Forcepoint X-Labs "10 IPI Payloads Caught in the Wild" (no direct relation to the three AI Frontier vendors; stored in the main brief's AI Security section)
  • OpenAI C2PA Conforming Generator → archived at security/2026-05-23-openai-c2pa-conforming.md
  • Project Zero Pixel 10 0-click → archived at security/2026-05-23-google-projectzero-pixel10.md
  • Mandiant PROMPTFLUX / PROMPTSTEAL → archived at security/2026-05-23-google-mandiant-promptflux.md
  • Mandiant GRIDTIDE update → same topic as yesterday's archive (5/22 google-ai-zeroday-attribution); no separate file today

Vendor configuration: see the ai_frontier category in intel/sources.yaml.

← Prev
AI Frontier · May 22, 2026
Next →
AI Frontier · May 24, 2026

🛡 Mandiant / GTIG — PROMPTFLUX & PROMPTSTEAL malware families
They call the Gemini API at runtime to rewrite their own source code / dynamically generate exfil payloads. Offensive "LLM use" has moved beyond simple pretext generation; defensive EDR signal recommendations are in the main brief's threat-intel section.
Adversaries Leverage AI — Google Cloud Blog

  • 🛡 Mandiant / GTIG — GRIDTIDE / UNC2814 disruption update
    Latest numbers: 53 confirmed victim organizations across 42 countries at takedown.
    Disrupting GRIDTIDE