AI Frontier · May 24, 2026

Latest developments from OpenAI / Anthropic / Google DeepMind. Security-related items are additionally archived under intel/ai-frontier/security/.

OpenAI

• 🛡 C2PA + SynthID watermarking now applied across all image output (since 5/19) — every image generated by ChatGPT / Codex / the API carries both C2PA provenance metadata and an invisible SynthID watermark, resistant to screenshotting, compression, and format conversion. Public verification tool: openai.com/verify. A direct counter to the AI deepfake disinformation chain. OpenAI — Advancing content provenance · Resultsense report · winbuzzer — SynthID

• Deep update to ChatGPT Business / Codex commercial workflows — appshots, goal mode to GA, improved browser annotation + browser-use tooling, remote locked-screen computer use, shared plugin workspace reuse, plus new admin adoption analytics. Releasebot — OpenAI updates

• ChatGPT for Personal Finance (Plaid integration, US ChatGPT Pro preview, 12,000+ financial institutions) — rollout continues. TechCrunch

Anthropic

• 🛡 Project Glasswing second update (5/22) — 10,000+ high/critical vulnerabilities, ~50 partners, a committed $100M in Mythos usage credits + $4M in open-source security donations — Cloudflare reported ~2,000 bugs (400 of them high/critical), with a false-positive rate below human review; Mozilla used Mythos to identify and fix 271 vulnerabilities in Firefox 150. Anthropic plans to widen Glasswing's scope to "critical partners," including the US and allied governments. Anthropic — Glasswing second update · Benzinga report · IBM newsroom — Project Glasswing collaboration expands

• 🛡 Claude Security public beta — Claude Opus 4.7 patched 2,100+ vulnerabilities in three weeks; the newly released cyber-verification tooling is open to qualifying security teams (scanning, triage, fix generation). SiliconANGLE

• Claude integrated into the SAP Business AI Platform (SAP Sapphire, 5/12–5/13) — via the MCP protocol, Claude becomes the primary reasoning / agent capability across S/4HANA, SuccessFactors, Ariba, and third-party systems. Treasury Manager → CFO briefing use case. SAP News Center · ERP Today — Claude + Joule + MCP · Seoul Economic Daily — Autonomous Enterprise

• Claude for Small Business (5/13) — packaged QuickBooks / PayPal / HubSpot / Canva / DocuSign / Google Workspace / M365 connectors plus ready-to-use workflows. InfoWorld

• Billing changes (effective 6/15) — programmatic Claude usage is separated out from chat subscription allowances; Pro gains $20, Max 5x gains $100, Max 20x gains $200.

• Talent moves — Andrej Karpathy joins Anthropic (focused on pretraining research). May 2026.

• Aftermath of Code with Claude 2026 (5/6, San Francisco) — Managed Agents, proactive workflows, dreaming, multi-agent orchestration, outcomes, webhooks, and a public beta of self-hosted sandboxes. MIT Tech Review published an industry reflection: "coding's future, whether you like it or not"

Google DeepMind / AI

• Gemini 3.5 Flash released (Google I/O 2026, 5/19) — beats Gemini 3.1 Pro on coding / agent benchmarks, with output tokens/second 4x faster than other frontier models. Available in the Gemini app / Search AI Mode / Antigravity / Gemini API / Gemini Enterprise. Gemini 3.5 Pro is already in internal use, with a broad launch next month. Google Blog — Gemini 3.5 · CNBC

• Price changes — Gemini Ultra cut from $250/month to $200/month, with a new Developer tier at $100/month. Heygotrade — I/O 2026 recap

• 🛡 Co-Scientist — a multi-agent research system in Nature (5/19 paper) — a Gemini-driven multi-agent system that iteratively generates, debates, and evolves new hypotheses. It included an independent CBRN (chemical, biological, radiological, nuclear) misuse evaluation and custom safety classifiers flagging unethical research goals and unsafe output. Notable: this is empirical evidence of safety engineering for agentic systems reaching Nature. DeepMind — Co-Scientist

• 🛡 C2PA + SynthID detection natively integrated into Google Search + Chrome (I/O 2026, 5/19) — announced the same day as OpenAI's, in coordination; the start of platform-level provenance UX.

• Project Zero / Google Security — no new posts this week, but the 5/13 "Pixel 10 zero-click exploit chain" is still circulating; browser and Android red-team communities are pulling it as an RCA case study.

• DeepMind hires 20+ Contextual AI researchers — an $80–90M licensing deal; the hoarding of agent research talent continues.

• Frontier Safety Framework update — DeepMind introduces a Harmful Manipulation Critical Capability Level (CCL), tracking and evaluating a model's ability to purposefully and systematically alter beliefs / behavior. Gemini 3 Pro is already under this evaluation regime.

---

Security-related items (see the ai-frontier/security/ subdirectory)

Newly archived today (5 items):

• 2026-05-24-anthropic-glasswing-second-update.md— Glasswing second-round 10K+ vulnerability update
• 2026-05-24-openai-c2pa-synthid-rollout.md— OpenAI's 5/19 full rollout of content provenance
• 2026-05-24-google-deepmind-co-scientist-safety.md— the CBRN safety evaluation in the Co-Scientist Nature paper
• 2026-05-24-google-c2pa-synthid-search-chrome.md— C2PA / SynthID integration into Google Search + Chrome
• 2026-05-24-anthropic-claude-security-beta.md— three weeks of results from the Claude Security public beta

🛡 = security-related.