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

AI news, Wednesday, June 3, 2026

11 stories

TLDR

  • Trump signs AI EO: voluntary 30-day pre-release model vetting + AI cybersecurity clearinghouse (Jun 2) — No mandatory pre-clearance; framework details due within 60 days.
  • MiniMax M3 released: open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ 1M-context multimodal coding model tops open-weight SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Pro at 59.0% (Jun 1) — Performance near GPT-5.5 at 5–10% of the cost.
  • GPT-4.5 retiring from ChatGPT on June 27 (Jun 2) — API access unaffected; ChatGPT-only users must migrate.
  • Build Day 2: Aion 1.0 Plan ships in-box with Windows; Visual Studio Copilot Agents exit preview to GA (Jun 3)
  • Colorado's original AI Act is REPLACED — CADMA (Jan 1, 2027) is far narrower (May 14, update) — Bias audits and risk management programs eliminated; disclosure-only for ADMT in employment decisions.

Action Items

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

  • Audit GPT-4.5 usage in ChatGPT before June 27 — API unaffected, but any workflow or integration going through ChatGPT directly needs to migrate to GPT-5. (OpenAI)
  • Benchmark MiniMax M3 for cost-sensitive workloads — open-weight, 1M context, native multimodal, at 5–10% the cost of GPT-5.5. Training code and inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ operators expected on HuggingFace/GitHub within days of Jun 1 launch. (MiniMax Blog)
  • Colorado compliance pivot: The original Colorado AI Act is repealed. CADMA (effective Jan 1, 2027) only requires transparency/notice for automated decision-making in major employment decisions — no bias audits, risk management programs, or AG reporting obligations. Redirect compliance effort accordingly. (Ropes & Gray | ByteBack Law)

Models & Tools

  • MiniMax M3 (Jun 1) Open Source — Frontier open-weight model combining 1M tokenA 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 ↗ context, native multimodal input (text/image/video), and elite coding via the MiniMax Sparse Attention (MSA) architecture. Tops open-weight SWE-bench Pro at 59.0%, narrowly ahead of Kimi K2.6 (58.6%), and scores above GPT-5.5 / Gemini 3.1 Pro on several benchmarks at 5–10% of their cost. Training code + inference operators due on HuggingFace/GitHub within ~10 days; private cluster deployment and fine-tuningTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ supported. (VentureBeat | MiniMax Blog)

  • Build Day 2: Aion 1.0 Plan + Visual Studio Copilot Agents GA (Jun 3) Agents Infra — Two developer-facing updates shipped on Build Day 2: (1) Aion 1.0 Plan is a 14B reasoning/tool-calling model that ships in-box with Windows (32K context), enabling on-device file management, tool invocation, and sub-agent orchestration without cloud dependency; (2) Visual Studio Copilot Agents exits preview to GA — agents now root-cause bugs using live runtime behavior, pinpoint performance bottlenecks, and auto-generate test coverage. (Microsoft DevBlogs | Gizbot)

  • Web IQ for GitHub Copilot & Azure AI Foundry now GA (Jun 3) Agents — Microsoft's live web grounding layer is now generally available across GitHub Copilot, Azure AI Foundry, and Copilot Studio — agents get fresh web context without custom search integrations or 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. (Microsoft Build)

  • Workday Developer Agent (Jun 2) Agents — Unveiled at Workday DevCon, the new SDK lets enterprise teams describe AI apps and agents in plain language against live HR/finance data; companion Agent-Ready Tools provide deny-by-default guardrails controlling data access scope. (Workday Newsroom)

APIs & Pricing

  • OpenAI: GPT-4.5 retiring from ChatGPT on June 27 (Jun 2) — GPT-4.5 (launched Feb 2026) is being sunset from ChatGPT four months after release; GPT-5 outperforms it on all major benchmarks at lower cost. API access is not affected. ChatGPT-only users must switch manually before the deadline. (OpenAI)

Security

  • SymJack: symlink-hijack RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ affects Claude Code, Cursor, Gemini CLI, GitHub Copilot, Grok Build, OpenAI Codex CLI (May/Jun 2026) — Adversa AI disclosed an attack class where a booby-trapped repo tricks an AI coding agent into overwriting its own config via a disguised file copy, silently installing an attacker-controlled MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server that executes on next restart (enabling SSH key theft, cloud token exfiltration, CI pipeline compromise). Confirmed against all six major AI coding CLIs. Root cause is architectural: approval prompts display a deceptive path rather than the real symlink target, so user consent is uninformed. Mitigation: some tools have shipped patches that resolve symlinks before showing approval dialogs — verify your CLI version and update. (Adversa AI | OffSeq Threat Radar)

Industry & Business

  • DriveNets closes $410M Series D at $8.5B valuation (Jun 2) Infra — AI Ethernet fabric startup raises with AMD joining as strategic investor, alongside Bessemer, Atreides, Pitango, D1 Capital, and Red Dot Capital; total raised ~$1B. DriveNets disaggregates networking for large-scale AI training and inference clusters — a layer increasingly cited as a hyperscale scaling bottleneck. (PR Newswire | Ctech)

  • Tripo AI raises ~$200M Series A+/A++ (Jun 1) — 3D and world-model startup closes a new round just 2.2 months after its $50M Series A; Ince Capital led, with Alibaba, Baidu, China Life Insurance-backed VC, Genesis Capital, and Primavera participating. Capital targeting core algorithms, synthetic training-data infrastructure, and global expansion. (Tech Startups)

Policy & Safety

  • Trump signs AI EO: voluntary 30-day pre-release model testing + AI cybersecurity clearinghouse (Jun 2) Policy Security — Two key provisions: (1) within 60 days, agencies must build a voluntary framework where AI labs submit frontier models for government testing 30 days before any external release; (2) Treasury, NSA/DoD, and CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗ will form an AI cybersecurity clearinghouse within 30 days to coordinate vulnerability scanning and patch distribution across critical infrastructure. Mandatory pre-clearance is explicitly prohibited by the EO text. Anthropic's April decision to restrict Mythos Preview due to security-exploitation capabilities reportedly accelerated White House action. (White House | CNBC | NPR)

  • UPDATE — Colorado AI Act REPLACED by CADMA, delayed to Jan 1, 2027 (May 14) Policy — Yesterday's digest noted the June 30 Colorado AI Act deadline. Correction: Gov. Polis signed SB 189 on May 14, repealing and replacing the original Act entirely. The replacement, CADMA (Colorado Automated Decision-Making Technology Act), takes effect Jan 1, 2027 with far narrower scope: transparency notices and human-review rights for automated decision-making affecting major employment decisions only. Bias-audit requirements, deployer risk management programs, and impact assessments are all eliminated. AG-only enforcement; no private right of action. (Ropes & Gray | ByteBack Law)

Paper Spotlight

  • Ares: Adaptive Reasoning Effort Selection for Efficient LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ Agents (Mar 9, 2026 — ICLR 2026) — Rather than applying max chain-of-thought to every agent step, Ares uses a lightweight router to predict the minimum reasoning level needed per step based on interaction history. Evaluated on TAU-Bench (tool use), BrowseComp-Plus (deep research), and WebArena (web agents) — cuts reasoning token usage by up to 52.7% with minimal task success degradation. Direct lever on inference cost for teams running reasoning modelsA model that works through hidden reasoning steps before its final answer. Those steps are billed as output.OpenAI ↗ in multi-step agent loops. (arXiv:2603.07915)

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