🎯 Top 3 Things to Know
1. Anthropic is closing a $30 billion round at over $900 billion valuation, vaulting past OpenAI as the most valuable private AI company. The round is expected to sign this week, co-led by Sequoia, Dragoneer, Altimeter, and Greenoaks at roughly $2 billion each, with Founders Fund and General Catalyst joining. It is the second $30 billion raise in 14 weeks. Anthropic told investors to expect $10.9 billion in Q2 revenue, more than double the prior quarter. The new mark clears OpenAI's $852 billion March valuation, giving frontier-model buyers a sharper signal that supplier concentration is shifting rather than narrowing. The line to watch in the term sheet is the customer commit disclosure: recent enterprise wins (KPMG 276K seats, PwC 30K seats, SAP, EY-Microsoft) are revenue-bearing through 2027, and that is the math outside investors will scrutinize first. Bloomberg
2. Anthropic disclosed that Claude Mythos has identified over 10,000 critical software vulnerabilities, and pushed Claude Security to public beta for Enterprise customers. Mythos Preview scanned more than 1,000 critical open-source projects, surfaced 23,019 candidate findings, and external triage of the sample landed at a 90.8% true-positive rate. As of May 22, Anthropic has formally disclosed 1,596 vulnerabilities across 281 projects, with 97 patched. One finding in wolfSSL, a cryptography library shipped on billions of devices, would have let an attacker forge bank or email certificates. The customer-facing piece is Claude Security, which patched 2,100 vulnerabilities during a three-week pilot. The bottleneck has shifted from discovery to disclosure throughput. AI can scan faster than maintainers can patch, which puts the 90-day coordinated-disclosure clock under real pressure. Anthropic's CVD dashboard is the place to watch the gap. Anthropic
3. Microsoft made computer-using agents generally available in Copilot Studio, alongside the GA of Microsoft 365 E7 and Agent 365. The agents drive websites and desktop apps through the UI rather than scripted selectors, which removes the brittleness that kept RPA stuck for a decade. The bigger story is Agent 365 at $15 per user per month: a single control plane to observe, govern, and secure agents across Copilot Studio, AWS Bedrock, and Gemini imports, with Defender and Intune controls extending to local Windows agents. This is the first credible enterprise answer to agent sprawl. Anyone shipping their own agent platform now has to defend the build-versus-buy math against a vendor that already owns the identity, endpoint, and tenant boundary. Microsoft Copilot Blog
🚀 Frontier Models & Features
- Anthropic appoints KiYoung Choi as Korea Representative Director ahead of a Seoul office opening. Choi joins from Snowflake. Anthropic's Economic Index shows Koreans use Claude at 3.5x the rate expected for population size. SK Telecom and Law&Company are named anchor customers. Anthropic
- OpenAI shipped three realtime voice API models (GPT-Realtime-2, GPT-Realtime-Translate, GPT-Realtime-Whisper) plus GPT-5.5-Cyber, the defensive analog to Claude Security. OpenAI
🔬 Research Worth Reading
Is Agent Memory a Database? Rethinking Data Foundations for Long-Term AI Agent Memory (Orogat & Mansour / Concordia University). arXiv
- TL;DR: Reframes long-term agent memory as a data-management workload where correctness depends on the trajectory of state, not individual records. Formalizes it as Governed Evolving Memory (GEM) with four state-level operators (ingest, revise, forget, retrieve) instead of CRUD.
- Stat: Replays of public agent traces show roughly 40% of stored "facts" go stale within ten subsequent sessions, which conventional vector stores have no operator to revise.
- Apply it: Audit your agent's memory writes by intent (ingest vs. revise vs. forget). If your current store can only append, the failure mode is data foundations, not the retriever.
From Model Scaling to System Scaling: Scaling the Harness in Agentic AI (Gu, Al-Tawaha, Niu, Jia & Jin / UC Berkeley and Virginia Tech). arXiv
- TL;DR: Treats the harness around a foundation model as the first-class scaling object once capability plateaus, with three core bottlenecks: context governance, trustworthy memory, and dynamic skill routing. Ships CheetahClaws, a Python-native reference harness, and benchmarks it against Claude Code and OpenClaw.
- Stat: The same backbone model varies by more than 2x on long-horizon task success depending on the orchestration layer alone, holding tokens and model fixed.
- Apply it: Before swapping models for a "smarter" one, run the swap as a harness experiment. Hold the model constant and vary context-governance and routing. The 2x spread says most of the headroom is there.
The Range Shrinks, the Threat Remains: Re-evaluating LLM Package Hallucinations on the 2026 Frontier-Model Cohort (Churilov / Independent). arXiv
- TL;DR: Replicates earlier slopsquatting work on the current frontier cohort (Claude Sonnet 4.6, Claude Haiku 4.5, GPT-5.4-mini, Gemini 2.5 Pro, DeepSeek V3.2) across 199,845 Python and JavaScript prompts validated against PyPI and npm.
- Stat: Per-model rates compress to a 4.62% (Claude Haiku 4.5) to 6.10% (GPT-5.4-mini) band, down from 5.2% to 21.7% in 2024 work. The threat surface narrowed but did not close.
- Apply it: Keep the install-time package allow-list. A 5% floor still registers thousands of bad installs across a CI fleet, and a pre-install reputation check on model-suggested dependencies is still the cheapest mitigation.
🏢 Enterprise in the Wild
The Project Glasswing partner roster is itself a who-is-on-board snapshot: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks all have Mythos access for internal vulnerability work. The notable absence is any cloud-resident financial-data platform, suggesting the next partnership wave targets the financial-services compliance perimeter. CSO Online
🛠️ Tooling & Ecosystem
- Copilot Studio also shipped a redesigned workflows experience with a visual orchestration designer that replaces the previous tool-stitching pattern, plus Work IQ extensibility points for third-party agent context. Microsoft Copilot Blog
⚖️ Policy & Regulation
California Governor Gavin Newsom signed Executive Order N-6-26 on May 21, directing state agencies to study severance, subsidized employment, stock compensation, and cooperative ownership for workers displaced by AI. It is the first US executive action to treat AI displacement as a structural policy problem rather than a re-skilling program. Expect downstream procurement and disclosure requirements within 12 months. California Governor's Office
📌 Watch List
- AI vulnerability discovery vs. patch throughput: disclosure backlog is the new bottleneck.
- Agent governance as a control plane: Agent 365 sets the bar; rivals will follow.
- Harness as the scaling variable: a second wave of papers framing orchestration as the headroom.
- Long-term agent memory: state-trajectory framing is replacing record-level CRUD.
- State-level AI workforce policy: California's EO will pull other governors toward similar moves.