Another packed week for software developers. Microsoft opened Build 2026 on June 2 in San Francisco with seven in-house MAI models, the agent-native GitHub Copilot app, and a clear message that it wants to lean less on OpenAI. The day before, GitHub Copilot switched every plan to token-based billing, and the developer reaction was loud. Anthropic filed a confidential S-1 for its IPO on June 1, OpenAI brought its frontier models and Codex to AWS the same day, and JetBrains open-sourced its 12 billion parameter Mellum2 model.

On the security side, the Miasma npm worm tore through Red Hat packages and then spread further using a sneaky binding.gyp trick. GitLab cut 14 percent of staff, Uber trimmed its HR team, and Oracle wrapped up around 30,000 job cuts. Cursor shipped 3.7, Elixir reached 1.20, Angular hit v22, and Anthropic mapped a full year of AI-enabled cyberattacks to the MITRE framework. Here is everything that mattered.

Top Stories This Week

Microsoft Build 2026 Opens With Seven In-House MAI Models -

On June 2, Microsoft opened Build 2026 in San Francisco and announced a family of seven in-house MAI models. The two that matter most for developers are MAI-Code-1-Flash, Microsoft’s first coding model and an inference-efficient option built into Copilot, and MAI-Thinking-1, its first reasoning model. Microsoft AI CEO Mustafa Suleyman said the models can run on Azure to avoid paying third parties, and that after tuning for the consulting firm McKinsey, Microsoft beat GPT-5.5 with roughly ten times better cost efficiency.

The models go beyond Copilot. Microsoft said they will be available to developers on OpenRouter, Fireworks, and Baseten, with the ability to tune the weights directly for the first time, and it announced a frontier healthcare model built with Mayo Clinic. Microsoft framed all of this as the start of a superintelligence lab built on clean data trained from scratch, without distillation from other companies’ models.

The rest of the keynote leaned hard into agents. Microsoft launched Scout, an always-on work agent built on the open-source OpenClaw framework, two on-device Windows models called Aion 1.0 Instruct and Aion 1.0 Plan, Project Solara as an Android-based OS for agent devices, the Agent 365 control plane for managing agents, Work IQ APIs for grounding agents in workplace context, and the Majorana 2 quantum chip.

GitHub Copilot Token Billing Goes Live to Heavy Backlash -

On June 1, GitHub Copilot switched every plan from flat pricing to token-based billing under a new system called AI Credits. The old Premium Request Units model used to downgrade you to a cheaper model once you hit your limit, so you kept getting answers. That fallback is gone. Now every chat, code review, agent action, and workspace query draws down credits based on the tokens actually processed, and once your allotment is spent you either stop or pay for more.

The reaction was immediate. The phrase “what a joke” spread across Reddit, Hacker News, and X, with one developer reporting a projected jump from 29 dollars to nearly 750 dollars a month. Defenders pushed back too, noting that careful prompting keeps costs flat or lower and that the big bills come from heavy, iterative vibe coding. The timing was awkward, landing one day before Microsoft used Build to pitch its cheaper MAI coding model.

GitHub Ships the Agent-Native Copilot App and Opens the Copilot SDK -

Also on June 2 at Build, GitHub introduced the GitHub Copilot app, an agent-native desktop experience now in technical preview for Copilot Pro, Pro+, Business, and Enterprise users. A single My Work view shows active sessions, issues, pull requests, and background automations across connected repositories. Every session runs in its own Git worktree so parallel agents do not step on each other, and a new canvas surface lets you inspect, steer, and redirect agent work as it happens. The app runs on Windows, macOS, and Linux and supports cloud and local sandboxes.

GitHub also made the Copilot SDK generally available across Node.js and TypeScript, Python, Go, .NET, Rust, and Java, exposing the same agentic runtime that powers the app and CLI. The Copilot CLI was redesigned with a voice mode and an experimental tabbed mode for pull requests, issues, and gists from the terminal. Two days later, on June 4, GitHub added an Agent tasks REST API in public preview for Pro, Pro+, and Max users, so you can start and track Copilot cloud agent tasks from a script and fan out refactors across many repositories.

Anthropic Files a Confidential S-1 for Its IPO -

On June 1, Anthropic confidentially submitted a draft registration statement on Form S-1 to the SEC for a proposed initial public offering. The filing landed days after the company closed a 65 billion dollar Series H at a 965 billion dollar valuation. The number of shares and the price were not set, and the company said any listing depends on market conditions and the SEC review.

On June 3, Bloomberg reported that Anthropic picked Goldman Sachs, Morgan Stanley, and JPMorgan as lead underwriters. Reporting points to a possible listing as early as October 2026, which would put Anthropic ahead of OpenAI, which is preparing its own confidential filing. As The Next Web noted, Anthropic, OpenAI, and SpaceX all heading toward the public markets in the same window would be the largest cluster of tech IPOs since the dot-com era.

OpenAI Brings Frontier Models and Codex to AWS -

On June 1, OpenAI made its frontier models and Codex generally available on AWS. The pitch to enterprises is simple: bring OpenAI capabilities into AWS environments through the security, compliance, procurement, and billing workflows their teams already trust, which removes a big barrier to getting AI into production. OpenAI also said future availability on AWS will include Daybreak, its security-focused effort that bundles cyber models and Codex Security for secure code review, threat modeling, patch validation, and dependency risk analysis inside the everyday development loop.

JetBrains Open-Sources the Mellum2 Coding Model -

On June 1, JetBrains open-sourced Mellum2, a 12 billion parameter model built for the plumbing of agentic systems: routing, retrieval pipelines, sub-agent tasks, and private on-premises deployment. It uses a Mixture-of-Experts design with only 2.5 billion parameters active per token, which JetBrains says cuts inference time to less than half that of comparable dense models. Built on the Qwen3-MoE recipe with 64 experts, it ships in base, instruct, and thinking variants under the Apache 2.0 license on Hugging Face.

The point, as The New Stack put it, is what Mellum2 does not need. Claude Code and Codex run locally but route inference through Anthropic and OpenAI APIs. Mellum2 arrives with open weights so teams can own and run that layer on their own infrastructure.

Miasma npm Worm Spreads Through Red Hat Packages -

On June 1, Microsoft Threat Intelligence flagged 32 maliciously modified npm packages across more than 90 versions under the @redhat-cloud-services scope. As Palo Alto’s Unit 42 explained, the root cause was a compromised Red Hat employee GitHub account used to push malicious commits through the legitimate CI/CD publishing workflow, so the trojanized packages carried authentic provenance signatures. A preinstall hook ran an obfuscated dropper that pulled the Bun runtime and harvested credentials from GitHub, npm, AWS, Azure, GCP, HashiCorp Vault, and Kubernetes. The campaign stamps the marker “Miasma: The Spreading Blight” on attacker repositories and descends from the Shai-Hulud worm family.

On June 3, the campaign got worse. StepSecurity reported 57 packages across more than 286 versions compromised in under two hours, led by @vapi-ai/server-sdk with over 408,000 monthly downloads and the jagreehal package family including ai-sdk-ollama. This wave used a technique StepSecurity named Phantom Gyp. Instead of the preinstall or postinstall scripts that security tools watch, the attacker abused the binding.gyp file to trigger code through node-gyp during npm install. The worm self-propagates by republishing packages from any maintainer account it reaches. If you installed any affected version, treat the environment as compromised, rotate every exposed credential, and use npm ci --ignore-scripts in CI.

GitLab Cuts 14% of Staff in AI Restructuring -

On June 3, GitLab laid off about 350 employees, roughly 14 percent of its workforce, as part of a restructuring it detailed in May. The company is exiting 22 countries, flattening management layers, and reorganizing R&D into smaller teams as it pushes its Duo Agent Platform. GitLab reported first-quarter revenue of 264 million dollars, up 23 percent year on year, with 88 percent gross margins, and expects 30 to 35 million dollars in restructuring costs. It is the same pattern seen at Intuit, Amazon, Meta, Microsoft, and Oracle this year: record revenue alongside a shrinking headcount, with AI cited as both the reason for the growth and the justification for the cuts.

Developer Tools & Platforms

Cursor 3.7 Adds Canvas Design Mode and SDK Custom Tools

On June 4 and 5, Cursor shipped 3.7 with Design Mode for canvases, which lets you select and annotate UI elements directly to guide edits instead of describing them in text, plus an interactive context usage report that breaks down where tokens go across the system prompt, tool definitions, rules, and skills. Shared canvases can now open full-screen in the browser for presenting to a team.

On June 4, Cursor also shipped a batch of SDK updates across the TypeScript and Python SDKs. You can now choose how agent and run metadata is persisted, expose your own functions to the agent as custom tools, route local tool calls through Auto-review, and nest subagents to any depth. The release also brings reliability and platform fixes for running local and cloud SDK agents in CI and custom scripts.

OpenAI Ships the Dreaming Memory System for ChatGPT -

On June 4, OpenAI began rolling out a new memory system for ChatGPT called Dreaming. The update is built to fix staleness, correctness, and scalability problems that show up when memory has to work across hundreds of millions of users and multi-year histories. OpenAI calls the current version Dreaming V3 and says recent work cut the compute needed to serve it to free users by about five times. It is live for Plus and Pro users in the US first, with free and Go users and more countries to follow, and it doubles memory capacity for Plus and Pro.

OpenAI Updates GPT-Rosalind for Life Sciences -

On June 3, OpenAI updated GPT-Rosalind, its model built for life sciences research. The update folds in GPT-5.5’s agentic coding and tool use and adds two Codex plugins, Life Sciences Research and Life Sciences NGS Analysis, that connect the model to more than 50 scientific databases. OpenAI says the model leads GPT-5.5, Grok 4.3, and Gemini 3.1 Pro on its LifeSciBench evaluation while using 31 percent fewer tokens than GPT-5.5 on genomics workloads. The model is in research preview to eligible organizations worldwide through a trusted-access program.

Anthropic Deprecates Claude Opus 4.1 -

On June 5, Anthropic notified developers that Claude Opus 4.1 is deprecated, with retirement on the Claude API scheduled for August 5, 2026. The model name is claude-opus-4-1-20250805, and Anthropic recommends migrating to Claude Opus 4.8. If you are still pinned to 4.1 in production, this is the cue to plan the move.

Core Language and Framework Releases

  • Elixir 1.20.0: Released on June 3, Elixir 1.20.0 landed with broader type inference across clauses and applications, occurrence typing on case, cond, and with, warnings on redundant clauses, and new helpers like Integer.ceil_div/2 and Integer.popcount/1. It requires Erlang/OTP 27 or later. See the official announcement for the full list.
  • Angular v22: Announced on June 3 with a release event on June 5, Angular v22 promotes Signal Forms, Angular Aria, and asynchronous reactivity APIs toward stable. Webpack support and the @angular-devkit/build-angular builders are deprecated as the team focuses on TSGo support in the application builder.
  • Go 1.26.4: Released on June 2, Go 1.26.4 is a minor revision with security fixes to the crypto/x509, mime, and net/textproto packages, plus bug fixes to crypto/fips140, go/types, and os.

Security

Anthropic Maps a Year of AI-Enabled Cyber Threats to MITRE ATT&CK -

On June 3, Anthropic published an analysis of 832 accounts banned for malicious cyber activity between March 2025 and March 2026, mapping 13,873 actions across 482 techniques onto the MITRE ATT&CK framework. The headline finding: the share of medium and high risk actors using AI rose from 33 percent in the first half of the period to 56 percent in the second. About 67 percent of analyzed actors used AI for malware development.

As Help Net Security summarized, the deeper point is that AI is breaking the old link between an attacker’s skill and the danger they pose, because agentic systems can chain together lateral movement, credential dumping, and other late-stage steps that used to require real expertise. Anthropic says behaviors like autonomous killchain orchestration and real-time pivot decisions are not yet captured by MITRE ATT&CK, and it is in talks with MITRE about how the framework should evolve. The company also said the findings informed safeguards now deployed on its most capable models.

Google Patches an Exploited Android Zero-Day in 124-Flaw June Update -

On June 1, Google’s June 2026 Android security bulletin patched 124 vulnerabilities, including CVE-2025-48595, an integer overflow in the Android Framework that is under limited, targeted exploitation. The flaw lets a local attacker escalate privileges without user interaction, and it affects Android 14, 15, 16, and 16 QPR2. Google shipped two patch levels, 2026-06-01 and 2026-06-05, with the later one bundling kernel and chipset fixes from Qualcomm, MediaTek, and others. On June 2, CISA added CVE-2025-48595 to its Known Exploited Vulnerabilities catalog and gave federal agencies until June 5 to patch.

Anthropic Extends Project Glasswing to 150 Organizations -

On June 2, Anthropic extended Project Glasswing to roughly 150 new organizations across more than fifteen countries. The program gives security teams access to Claude for vulnerability discovery and defensive work, and Anthropic said it plans to share vulnerability-finding tools with trusted security teams to strengthen defenses. It pairs naturally with the MITRE report above as part of Anthropic’s wider cyber push.

Funding & Industry Deals

Anthropic Files for Its IPO and Names Underwriters -

On June 1, Anthropic filed its confidential draft S-1 with the SEC, and on June 3 it named Goldman Sachs, Morgan Stanley, and JPMorgan as lead underwriters. The move sets up a potential public listing as soon as October 2026, with run-rate revenue reported around 47 billion dollars at the time of its Series H.

Uber Cuts 23% of Its People and Places Team -

On June 3, Uber cut 23 percent of its People and Places division, the unit covering HR, recruitment, and workplace facilities. The affected roles are less than 1 percent of Uber’s 34,000 corporate staff, and the company said the move was about organizational overlap, not AI. In the same reporting, Uber said nearly 95 percent of its engineers use AI tools monthly and close to 70 percent of committed code is now AI-generated, after it burned through its entire 2026 AI coding budget in four months. Tracker TrueUp put 2026 tech layoffs at 149,935 across 363 events by that day.

Oracle Wraps Up Around 30,000 Job Cuts by June 15 -

Oracle is in the final stage of its largest workforce reduction, with thousands of employees reaching their separation dates between June 1 and June 15. The total is around 30,000 people, roughly 18 percent of its global workforce, with the deepest cuts in Oracle Health, the division built from the 28.3 billion dollar Cerner acquisition. The cuts come during strong financial results, as Oracle redirects resources toward AI infrastructure.

The Numbers That Matter

  • 7 In-house MAI models Microsoft launched at Build 2026, including its first reasoning and coding models
  • $965 Billion Valuation Anthropic carries into its confidential IPO filing
  • 12 Billion Total parameters in JetBrains Mellum2, with only 2.5 billion active per token
  • 32 Red Hat scope npm packages hit by the Miasma worm on June 1, before a 57-package wave on June 3
  • 124 Vulnerabilities patched in Google’s June Android bulletin, including one exploited zero-day
  • 56% Share of medium and high risk actors using AI in the second half of Anthropic’s study, up from 33%
  • 350 GitLab employees laid off, about 14 percent of the company
  • 23% Of Uber’s People and Places team cut on June 3
  • 30,000 Oracle jobs being eliminated by June 15
  • 6 Languages now supported by the generally available GitHub Copilot SDK

Quick Hits

  • GitHub Copilot Token Billing - June 1. All plans move to AI Credits token-based pricing, sparking heavy developer backlash.
  • Anthropic Confidential S-1 - June 1. Draft IPO filing submitted to the SEC, with underwriters named on June 3.
  • OpenAI Models and Codex on AWS - June 1. Frontier models and Codex generally available to AWS customers.
  • JetBrains Mellum2 - June 1. 12B Mixture-of-Experts coding model open-sourced under Apache 2.0 on Hugging Face.
  • Miasma npm Worm (Red Hat) - June 1. 32 @redhat-cloud-services packages trojanized through a compromised CI/CD pipeline.
  • Microsoft Build 2026 - June 2. Seven MAI models, Scout, Aion, Project Solara, Agent 365, Work IQ, and Majorana 2.
  • GitHub Copilot App and SDK - June 2. Agent-native desktop app in technical preview and the Copilot SDK goes GA.
  • Anthropic Project Glasswing - June 2. Extended to roughly 150 organizations in more than fifteen countries.
  • Android June Bulletin - June 2. CISA adds the exploited CVE-2025-48595 to its KEV catalog, deadline June 5.
  • Go 1.26.4 - June 2. Minor release with security fixes to crypto/x509, mime, and net/textproto.
  • GitLab Layoffs - June 3. About 350 jobs cut, 22 countries exited, focus shifted to the Duo Agent Platform.
  • Uber HR Cuts - June 3. 23 percent of the People and Places division eliminated.
  • Anthropic MITRE Report - June 3. A year of AI-enabled cyber threats mapped to MITRE ATT&CK.
  • GPT-Rosalind Update - June 3. New life sciences capabilities and two Codex plugins.
  • Elixir 1.20.0 - June 3. Broader type inference and occurrence typing.
  • Angular v22 - June 3. Signal Forms, Angular Aria, and async reactivity move toward stable.
  • OpenAI Dreaming Memory - June 4. New ChatGPT memory system rolling out to Plus and Pro in the US.
  • GitHub Agent Tasks REST API - June 4. Start and track Copilot cloud agent tasks from a script.
  • Cursor 3.7 and SDK Update - June 4 to 5. Canvas Design Mode, context usage report, and SDK custom tools with Auto-review.
  • Anthropic Opus 4.1 Deprecation - June 5. API retirement set for August 5, 2026, migrate to Opus 4.8.
  • Oracle Cuts - June 1 to 15. Around 30,000 jobs eliminated, deepest in Oracle Health.

The theme this week was control. Microsoft used Build to show it wants its own models, its own agents, and its own desktop surface for development work, while GitHub’s new billing made the cost of someone else’s models impossible to ignore. Anthropic moved toward the public markets, OpenAI planted its models inside AWS, and JetBrains made the case that some of the AI stack belongs on your own infrastructure. Meanwhile the Miasma worm and Anthropic’s MITRE report were a reminder that the same agentic tools shipping in product keynotes are already in the hands of attackers. Next week brings Apple’s WWDC and the iOS 27 reveal on June 8, with the rest of June still pointing toward Gemini 3.5 Pro. See you then.