Tuesday, June 2, 2026

TLDR

Microsoft Build 2026 keynote today: Windows Agent Framework MIT-licensed, Project Polaris replaces OpenAI models in Copilot by August, Windows Local AI runtime ships June 9 (Jun 2) — The full agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ stack is now confirmed and shipping.
Anthropic files confidential S-1 with the SEC (Jun 1) — $47B ARR, $965B valuation; no shares or price set yet but the IPO race is officially on.
Langroid CVE-2026-25879 (CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 9.8 Critical): SQLChatAgent RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ via prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ (Jun 1) — If agents have DB exec privileges, attackers can achieve RCE. Upgrade to v0.63.0 now.
Azure multi-day power outage (May 29–Jun 1) takes down Copilot, Azure ML, AI Search — Thunderstorm cascade-failed UPS and cooling at Azure East US; post-mortem due Jun 6.
Anthropic expands Project Glasswing to ~200 orgs (Jun 2) — Claude Mythos access now spans power, water, healthcare, comms, and hardware sectors; public Mythos-class release "coming weeks."

Things your engineering team should look at or do this week.

Upgrade Langroid to v0.63.0 immediately if you use SQLChatAgent. CVE-2026-25879 (CVSS 9.8) allows prompt injection to coerce LLM-generated SQL into RCE when the DB role has exec privileges (PostgreSQL pg_execute_server_program, MySQL FILE, MSSQL xp_cmdshell). (TheHackerWire)
Evaluate Windows Agent Framework and Windows Local AI for on-device agent workloads — the MIT-licensed WAF is your path to NPU-native agents without cloud dependency; the runtime ships in Windows 11 KB5039239 on June 9. (Microsoft Build)
Read the Azure post-mortem when it drops ~June 6 — the power-failure cascade that took down Azure ML and M365 Copilot for 3+ days will contain concrete lessons on AI infra resilience and redundancy design. (Azure Status)
If GitHub Copilot is part of your stack, start testing Project Polaris now — Microsoft's in-house model replaces GPT-4 as the Copilot default in August; opt-in fallback runs for three months. (ChatForest Build Recap)

Microsoft Build 2026: Full Agentic Stack Confirmed (Jun 2) Agents Infra — Satya Nadella's keynote shipped the announced roadmap: Windows Agent Framework v1.0 is MIT-licensed, Windows Local AI runtime (NPU-native, no cloud required) lands June 9 via KB5039239 for Snapdragon X Elite / Intel Lunar Lake / AMD XDNA PCs, Copilot Workspace exited beta to GA, and Office 365 Copilot Agent Mode becomes the default across Word/Excel/PowerPoint in a late-June rollout. (Microsoft Build | ChatForest Recap)
Project Polaris & MAI Models Announced (Jun 2) Agents — Microsoft unveiled its first in-house reasoning modelA model that works through hidden reasoning steps before its final answer. Those steps are billed as output.OpenAI ↗, MAI-Thinking-1 (not distilled from any external model), plus MAI-Image-2.5 and MAI-Image-2.5-Flash; Project Polaris — the coding-focused model — replaces GPT-4 Turbo as Copilot's default by August, with a three-month optional GPT-4 fallback window for teams that need it. (Windows News)
Anthropic expands Project Glasswing to ~200 orgs (Jun 2) Security Agents — Claude Mythos access extended by 150 new organizations across 15+ countries, now covering critical infrastructure sectors (power, water, healthcare, comms, hardware) not in the initial cohort. Partners have surfaced 10,000+ high/critical security flaws since launch; public Mythos-class access remains "coming weeks." (Anthropic Blog | CNBC)

CVE-2026-25879: Langroid SQLChatAgent — CVSS 9.8 Critical RCE (Jun 1) — SQLChatAgent prior to v0.63.0 executes LLM-generated SQL without sanitization. An attacker who controls agent input can inject prompts that coerce the model into generating SQL primitives (COPY … FROM PROGRAM, xp_cmdshell, MySQL FILE) achieving RCE on the database host. Fix: upgrade to v0.63.0, which defaults to a SELECT-only sqlglot-parsed allowlist; dangerous ops require explicit opt-in. (TheHackerWire | GitLab Advisory)

Anthropic files confidential S-1 with the SEC (Jun 1) Trending — Confidential draft registration filed for a potential IPO; $47B ARR (up from $10B last year), last-round valuation of $965B — topping OpenAI's $852B. No share count or price set; timeline depends on SEC review and market conditions. (CNBC | TechCrunch)
Tech layoffs paradox: 148K displaced in 2026, ML hiring up 59% (Jun 1) Hiring — AI is the stated cause of ~25% of 2026 tech cuts (the leading single reason for two consecutive months), and junior dev employment (ages 22–25) is down ~20% since 2024. Meanwhile ML engineering openings are 59% above the 2020 baseline — so if your team needs AI/ML or security talent, the current talent pool is deep. (TechTimes)

Azure East US power failure → 3-day Copilot/Azure ML outage (May 29–Jun 1) Infra — A thunderstorm front knocked out primary and backup UPS at Azure East US on May 29, cascading to cooling failure and forcing clusters offline; Azure ML, AI Search, and M365 Copilot were degraded or unreachable, peaking at 14K+ incident reports. Azure VP Mark Russinovich acknowledged a "rare synchronization failure in primary and backup power control systems." Recovery was ~70% by Jun 1; full post-mortem expected by June 6. Lesson: AI control planes co-located in a single AZ inherit physical infrastructure failure modes. (Windows News | Android Authority)

Colorado AI Act enforcement clock: 28 days out (Jun 2) Policy — Colorado's AI Act takes effect June 30, 2026, covering developers and deployers of high-risk AI systems making consequential decisions (education, employment, healthcare, housing, insurance, legal). Requires risk management programs, impact assessments, and anti-discrimination measures. If you ship to Colorado users, final review window is now. (Kiteworks)

Claude Mythos Preview leads SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Verified at 93.9% (as of Jun 1) — Anthropic's as-yet-unreleased model tops both SWE-bench Verified (93.9%) and SWE-bench Pro (77.8%), ahead of Claude Opus 4.8 (88.6%). The gap between the current best public model and the upcoming Mythos release is ~5 points on the hardest coding evals — a meaningful signal heading into Mythos's public launch. (SWE-bench)