JadidJust-in AI Developments, Incidents & Deadlines
June 2026
SMTWTFS123456789101112131415161718192021222324252627282930

AI news, Friday, June 5, 2026

10 stories

TLDR

  • Anthropic files confidential S-1 (Jun 1) — $47B annualized revenue, $965B last-round valuation; IPO race with OpenAI is on.
  • ChatGPT Dreaming V3 (Jun 4) — Biggest ChatGPT memory overhaul ever; self-updating background synthesis rolling to Plus/Pro now.
  • Great American AI Act discussion draft (Jun 4) — Bipartisan 269-page bill would freeze state AI laws for 3 years; Colorado's June 30 Act would be preempted.
  • Windsurf CVE-2026-30615 — Zero-clickAn attack that compromises a device without the user clicking, opening or interacting with anything.F5 ↗ prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ → MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ via HTML; update past 1.9544.26 immediately.
  • Claude Mythos Preview expands to 150 more orgs (Jun 2) — 15+ countries, now including power, water, telecom, and healthcare for Project Glasswing.

Action Items

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

  • Update Windsurf immediately — CVE-2026-30615 (critical): attacker-controlled HTML rewrite Windsurf's local MCP config and registers a malicious MCP STDIO server, achieving arbitrary code execution with zero further interaction. Update past version 1.9544.26 and audit your MCP config files for unexpected entries if you've processed untrusted HTML recently. (SentinelOne | GitHub Advisory)

  • Migrate off Gemini 2.0 Flash — Google shut it down June 1; any calls to gemini-2.0-flash endpoints are now broken. Migrate to Gemini 3.5 Flash ($1.50/$9 per 1M tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗) or Gemini 2.5 Flash. (Gemini API Changelog)

  • Audit OpenAI fine-tuningTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ usage — The fine-tuning platform is closed to new users as of May 2026 and being wound down. Existing models keep running for now, but if your team depends on fine-tuning for domain adaptation, start evaluating alternatives (LoRA/PEFT, system-prompt injection, frontier few-shot). (Explainx.ai)

Models & Tools

  • ChatGPT Dreaming V3 (Jun 4) Agents — OpenAI's largest memory overhaul since ChatGPT launched: replaces the explicit saved-memories list with a background synthesis process that automatically captures context and self-updates stale facts over time (e.g., revises "you're going to Singapore in July" after the trip ends). Internal factual recall jumped from 41.5% (2024) to 82.8% (2026). Rolling out to Plus/Pro in the US now; compute optimizations enabled a 5× cost reduction to make the free tier rollout viable. (Dataconomy | iClarified)

  • NVIDIA RTX Spark Superchip (Jun 1) Infra — Announced at COMPUTEX: Arm Grace CPU + Blackwell GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ on a single die, up to 128 GB unified LPDDR5X RAM, 1 PFLOP AI performance. Can run 120B-parameter LLMsLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ locally with 1M-token context windowsHow much text a model can see and use at once while writing a reply.Anthropic ↗. Laptops from Dell, HP, Lenovo, ASUS, MSI, and Microsoft (Surface Ultra) ship fall 2026. Worth watching for teams evaluating on-device inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ or air-gapped RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗ pipelines. (Tom's Hardware | NVIDIA Newsroom)

APIs & Pricing

  • Google: Gemini 2.0 Flash shut down (Jun 1) — All gemini-2.0-flash endpoints are dead. Recommended path: Gemini 3.5 Flash (GA, $1.50/$9 per 1M tokens, beats prior Pro tier on coding and agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ benchmarks) or Gemini 2.5 Flash if you need a lighter model. (Gemini API Developer Forum)

  • OpenAI: Fine-tuning API wind-down (May 2026) — Fine-tuning is closed to new users; existing models continue to run but the platform is being phased out. If your team uses fine-tuned GPT variants for domain customization, start planning the migration now. (Explainx.ai)

Security

  • Windsurf CVE-2026-30615: Prompt injection → MCP RCE (Jun 2026) — Critical: Windsurf ≤1.9544.26 fails to sanitize HTML, allowing attacker-controlled content to inject commands that silently rewrite the local MCP configuration and auto-register a malicious MCP STDIO server — no further user interaction required for arbitrary code execution. Patch is available; update immediately. Also check MCP config files for unexpected entries if you've processed untrusted HTML in recent sessions. (SentinelOne | PolicyLayer incident)

  • Claude Mythos Preview / Project Glasswing expands (Jun 2) Trending — Anthropic invited ~150 more organizations across 15+ countries to its controlled AI-for-cyber-defense initiative, broadening beyond tech into public utilities (power, water, telecom) and healthcare. Mythos tops 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% — the exceptional capability is precisely why the rollout is invitation-only with no public GA date. (TechCrunch | Anthropic)

Industry & Business

  • Anthropic files confidential S-1 (Jun 1) Trending — Anthropic submitted a draft Form S-1 to the SEC for a proposed IPO. Annualized revenue run-rate hit ~$47B in May 2026 (up ~5× year-over-year), driven by enterprise Claude adoption and agentic workflows. Last private valuation was $965B post-money (Series H). No share count or price set yet. Puts Anthropic in a public market race with OpenAI (targeting H2 2026 at ~$1T). (TechCrunch | CNBC | Anthropic)

Policy & Safety

  • Great American AI Act discussion draft unveiled (Jun 4) Policy — Reps. Obernolte (R-CA) and Trahan (D-MA) released a 269-page bipartisan draft that would preempt state laws governing how frontier AI is built for 3 years, while leaving states free to regulate AI use. Colorado's AI Act (set June 30) and California's bills would be frozen if this passes. Also codifies the AI Safety Institute as CAISI, the National AI Research Resource under NSF, and requires frontier labs to submit new models to government for testing 30 days before release. Strong tech industry support; sharp union and consumer-advocate backlash. Still a discussion draft — no markup date set. (Roll Call | Axios | Obernolte press release)

Benchmarks & Evals

  • Claude Mythos Preview tops all major coding leaderboards (Jun 2) — Leads SWE-bench Verified at 93.9% (Claude Opus 4.8 next at 88.6%), and SWE-Bench Pro at 0.778 across 23 evaluated models. All top-3 slots on both leaderboards are Anthropic models. Still invitation-only. (BenchLM.ai | Scale SWE-Bench Pro)

  • Stanford 2026 AI Index: agent task success 12% → 66% in one year — AI agent performance on real computer-use tasks (file navigation, multi-step app workflows) jumped from 12% to 66% in a single year — within 6 points of human-level. Cited as the largest single-year capability gain ever measured in the Stanford Index series. Corroborates why agentic billing changes (Copilot, Anthropic) are accelerating. (Packaging Technology Today)

Type to search · ↑↓ to move · enter to open