Another heavy week for software developers, and this time the frontier labs set the tone, with the US government’s hand visible on almost every release. On June 26, OpenAI previewed its GPT-5.6 family but shipped it as a limited, government-gated preview rather than a normal launch. Four days earlier, Google made its Interactions API the official front door to Gemini, and Anthropic dropped a report claiming Claude now writes more than 80 percent of its own code.

The rest of the week kept pace. An attacker poisoned the Leo Platform npm scope with a self-spreading worm, Cursor quietly bought open-source rival Continue and shut it down, OpenAI expanded its Daybreak security push, and DeepReinforce open-sourced a strong coding model family. Nx and GitKraken both launched tooling for the multi-agent workflow that is quickly becoming normal, Next.js pushed frameworks further toward agents, Envoy AI Gateway hit v1.0, and Baseten raised $1.5 billion. Here is everything that mattered.

Top Stories This Week

OpenAI Previews GPT-5.6 Sol Under a Government-Gated Rollout -

On June 26, OpenAI began a limited preview of the GPT-5.6 series: Sol, the new flagship, Terra, a balanced model that OpenAI says matches GPT-5.5 at roughly half the cost, and Luna, a fast and cheap model for high-volume work. Sol sets a new state of the art on Terminal-Bench 2.1, the benchmark for complex command-line workflows that need planning, iteration, and tool coordination, which is exactly the kind of work agentic coding tools lean on. On paper this is a straightforward generational step up from GPT-5.5.

The unusual part is how it shipped. Instead of a broad launch, OpenAI said it is starting with a preview for a small group of trusted partners through Codex and the API, at the request of the US government, with general availability planned for the coming weeks. The company said it previewed the models and their capabilities to the administration ahead of time as part of work on a cyber Executive Order framework. OpenAI made its discomfort plain, writing that it does not believe this kind of government access process should become the long-term default because it keeps the best tools from the developers, enterprises, and cyber defenders who need them. After Anthropic’s Fable 5 was pulled by an export directive earlier in the month, a pattern is forming: frontier model releases now pass through Washington first.

Google Makes the Interactions API the Front Door to Gemini -

On June 22, Google announced that the Interactions API has reached general availability and is now the primary interface for Gemini models and agents. It went into public beta in December 2025, and the GA release brings a stable schema plus managed agents, background execution, and combined tool use. The legacy generateContent endpoint still works, but Google was clear that frontier capabilities for long-running models and agents will land exclusively on Interactions, and that all of its documentation now defaults to the new API with a toggle to flip snippets back.

The design tells you where API thinking is heading in the agent era. As API Evangelist put it, the old request and response shape, where you resend the whole conversation every call and get a single completion, was never going to carry agents that think, call tools, and run for minutes at a time. The Interactions API replaces role-based messages with typed steps like user_input, function_call, and model_output, keeps state server-side, and can provision a remote Linux sandbox for an agent in a single call. This is the same direction as OpenAI’s Responses API and Anthropic’s tool platform, so the major labs are now converging on stateful, agent-native interfaces. If you build on Gemini, treat Interactions as the path forward and plan migrations off generateContent for anything new.

Anthropic Says Claude Writes More Than 80% of Its Code -

On June 22, the Anthropic Institute published a report called When AI Builds Itself, and the headline number got everyone’s attention: more than 80 percent of the code merged into Anthropic’s codebase in May was authored by Claude, up from low single digits before Claude Code launched in research preview in February 2025. As the report itself describes, the typical engineer now ships about 8 times as much code per quarter as in the 2021 to 2024 baseline, with humans directing and reviewing rather than typing. On open-ended engineering problems that start without clear specifications, Claude’s success rate climbed to 76 percent in May, a 50-point jump in six months.

Anthropic framed this as early evidence that AI is already accelerating AI development, a possible path toward recursive self-improvement, while being careful to say that full self-improvement has not arrived and is not guaranteed. For working developers, the more useful signal is cultural: Anthropic says designers and PMs now check in code too, and that the bottleneck has moved from writing code to reviewing it. If 8x more code is landing, then review, testing, and architecture become the scarce skills, not typing speed. Take the exact figures with a grain of salt since they come from the vendor selling the tools, but the direction is hard to argue with.

Leo Platform npm Scope Poisoned by a Self-Spreading Worm -

The supply chain attack of the week hit the Leo Platform JavaScript ecosystem. On June 24, an attacker used a compromised maintainer account named czirker to publish malicious versions of more than 20 Leo Platform npm packages in a coordinated burst of a few seconds. StepSecurity found the payload structurally identical to the Miasma campaign from earlier in June, sharing the same binding.gyp install hook and obfuscation chain, which points to the same threat actor moving to a new target. The worm steals secrets from GitHub Actions runners, cloud credential stores, package registries, and password managers, then exfiltrates them through the victim’s own GitHub token and republishes backdoored packages on its own to keep spreading.

The nastiest detail, as Endor Labs documented, was dist-tag manipulation. For leo-sdk, the attacker repointed the latest tag to a malicious build while the clean 7.x release stayed available only under a separate tag, so any project running npm install leo-sdk without an explicit version pin pulled the worm. This is the second worm-style npm compromise in a month after the Mastra attack, and the playbook is now familiar: one stolen credential, an install-time hook, and self-propagation through CI. If you installed any Leo Platform package on or after June 24, pin to a known-clean version, rotate every token and cloud credential exposed in your pipelines, and rebuild from clean lockfiles.

Cursor Quietly Acquires Open-Source Rival Continue -

On June 22, Cursor acquired Continue, the open-source AI coding assistant, in a quiet acqui-hire that shut the product down. Continue had built up 34,300 GitHub stars and 4,800 forks and raised around $5 million as an open alternative to GitHub Copilot. The team pushed a final 2.0.0 release that stripped telemetry and tidied the code as a deliberate handoff to the community, and under its Apache 2.0 license the codebase stays public for anyone to fork.

This is the latest beat in a steady acquisition drumbeat from Cursor maker Anysphere, which has previously picked up Supermaven and code review startup Graphite. Unlike Graphite, which kept operating, Continue looks shuttered, with co-founder Nate Sesti joining Cursor while other founding engineers went elsewhere. The move lands the same week SpaceX’s $60 billion deal for Anysphere is still working through, so the company is consolidating talent even as its own ownership shifts. For developers who liked Continue precisely because it was open and model-agnostic, the closure is one less independent option in an increasingly consolidated market.

DeepReinforce Open-Sources the Ornith-1.0 Coding Models -

On June 25, DeepReinforce released Ornith-1.0, an open-source model family built for agentic coding. The lineup spans four sizes, from a 9B dense model to a 397B mixture-of-experts flagship, all shipped under the MIT license on Hugging Face and post-trained on top of Gemma 4 and Qwen 3.5. The interesting twist is that instead of pairing a model with a fixed, human-designed harness, Ornith-1.0 learns to write its own scaffold during reinforcement learning, jointly optimizing the harness and the solution.

The 397B model scores 82.4 on SWE-Bench Verified and tops Claude Opus 4.7 on the headline benchmarks, though it trails Opus 4.8 and the larger GLM-5.2, so the state-of-the-art claim is scoped to open models of comparable size. The models target terminal-native coding agents and repository-scale work like multi-file refactors and test-driven patches, with the 9B build aimed at single-GPU and edge setups. For teams that want a capable coding model they can self-host without sending code to a closed API, this is a meaningful new option, and the MIT license means no strings attached.

Developer Tools & Platforms

Nx Launches Polygraph for Cross-Repo Agents -

On June 23, Nx launched Polygraph, a service that connects multiple repositories into a single synthetic monorepo that AI coding agents can work across as if it were one codebase. Nx’s founders, former Google Angular engineers, argue that agents do fine inside a single repo but tend to stall when a change spans repositories or depends on someone else’s work. Polygraph combines repos with context from agent traces to give agents portable, shared memory. Nx calls it a meta-harness because it sits around the coding agents rather than being one, wiring them together over the agent-to-agent (A2A) protocol. It supports Claude Code, Codex, and OpenCode today, works only with GitHub at launch, and is free during early access.

GitKraken Ships Code Flow and the Kepler Agent IDE -

Also on June 23, GitKraken introduced Code Flow, a framework for how work moves between developers, coding agents, repositories, reviews, and production. The pitch is that the systems around code production, from review to conflict resolution to commit hygiene, were not built for the volume and velocity of multiple agents running at once. To back it, GitKraken launched Kepler, a purpose-built agent development environment for Windows, Mac, and Linux, alongside GitKraken Desktop 12 with a new agent mode and GitLens 18 with agentic capabilities for the VS Code and Cursor ecosystem. The framing matches what a lot of developers now actually do: spin up a fleet of agents in one tool, drop into an IDE to hand-edit, then wrestle a merge conflict before shipping.

Next.js 16.3 Bets on Agent-Driven Development -

On June 25 and 26, Next.js 16.3 Preview arrived with the clearest sign yet that frameworks are now being designed for coding agents as a first-class user. The AI Improvements post ships version-matched docs delivered through AGENTS.md, first-party Skills for multi-step workflows, an agent-browser that can drive a real browser and introspect React component trees, and actionable errors with a “Copy as prompt” button. Alongside it, Instant Navigations adds opt-in rendering and prefetching primitives that make server-driven routing feel as snappy as a single-page app without dropping the server. Paired with Vite 8.1, which pushed the Rolldown-based bundler to 41.6 million weekly downloads this week, the direction is hard to miss: the frontend toolchain is being rebuilt around agents and the AI workflow.

Envoy AI Gateway and Kubernetes Security Profiles Operator Reach v1.0 -

Two notable open-source infrastructure projects hit their first stable release on June 23. Envoy AI Gateway 1.0 is the first GA of the CNCF Envoy Gateway extension built for AI traffic, with a now-stable control-plane API, a single OpenAI-compatible interface across 16 providers, a full Model Context Protocol gateway, and multi-tenant, quota-aware routing. It arrived after 16 months of work led by maintainers at Bloomberg, Nutanix, and Tetrate, and the same code already runs in production at those companies. If you are stitching together LLM providers behind a homegrown proxy, this is a credible vendor-neutral option built on the battle-tested Envoy data plane. Separately, the Kubernetes Security Profiles Operator reached v1.0, graduating all of its CRDs from alpha and beta to a stable v1 API for managing seccomp, SELinux, and AppArmor profiles, and shipping a new spoc command-line tool. The release followed a third-party security audit that turned up zero critical vulnerabilities.

Security

OpenAI Expands Daybreak With Patch the Planet and GPT-5.5-Cyber -

On June 22, OpenAI expanded its Daybreak cybersecurity program with Patch the Planet, an open-source patching initiative, an updated Codex Security plugin, and the full release of GPT-5.5-Cyber, its most capable defensive model. OpenAI’s argument is a real shift in framing: its models now find vulnerabilities faster than defenders can fix them, so the new bottleneck is patching rather than discovery. The company said Codex turned up a year-old flaw in the OpenBSD kernel and flagged patterns matching six dnsmasq vulnerabilities that later got CVE numbers. GPT-5.5-Cyber scored 85.6 percent on the CyberGym benchmark, up from 81.8 for standard GPT-5.5, with access restricted to vetted defenders through the Trusted Access for Cyber program. Launch partners for the new Daybreak Cyber Partner Program include Accenture, Cisco, CrowdStrike, IBM, Okta, Palo Alto Networks, and Wiz.

Cisco and PTC Zero-Days Hit CISA’s Catalog With a June 28 Deadline -

On June 25, CISA added two actively exploited flaws to its Known Exploited Vulnerabilities catalog with a June 28 deadline for federal agencies. The first is CVE-2026-20230, a critical server-side request forgery flaw in Cisco Unified Communications Manager that lets an unauthenticated remote attacker write files to the underlying OS and later escalate to root. Cisco patched it on June 3 and saw no exploitation at the time, but threat detection firm Defused observed it being used in attacks the following weekend. The second is CVE-2026-12569, a critical remote code execution bug in PTC Windchill and FlexPLM that exploits deserialization of untrusted data. Both fall under Binding Operational Directive 26-04, which has driven an unusually fast pace of KEV additions this month, so if you run either product, patch now rather than waiting on the directive’s clock.

Funding & Industry Deals

Baseten Raises $1.5 Billion for AI Inference -

On June 22, Baseten announced a $1.5 billion Series F led by Altimeter Capital, Conviction, and Spark Capital, with Sands Capital and Wellington Management as co-leads. The round came in two tranches at valuations of $13 billion and $11 billion, a steep jump from the $5 billion valuation of its $300 million Series E in January. It is Baseten’s fourth raise in 18 months, reflecting how much money is chasing the inference layer that actually runs models in production, with customers including Cursor and Clay. Separately, AI chipmaker Groq closed $650 million to scale its inference cloud, six months after Nvidia hired away its founder and key staff in an acqui-hire-style deal.

Superhuman Buys GPTZero and MoEngage Buys Aampe -

On June 23, Superhuman acquired AI detection startup GPTZero, the tool built by Princeton grad Edward Tian that grew to more than 19 million registered users and $30 million in annual recurring revenue on just $13.5 million raised. Superhuman, the company formed when Grammarly bought the Superhuman email app and rebranded, already had its own AI detector and reasoned that two are better than one. The same day, India’s MoEngage acquired San Francisco startup Aampe in an all-cash deal reportedly worth tens of millions, betting that assigning a dedicated AI agent to each customer will become the future of marketing. Around 20 Aampe employees join MoEngage, pushing its headcount to roughly 820.

Layoffs: Walmart, Cisco, and Amperity

The Numbers That Matter

  • 80%+ Share of Anthropic’s merged May code that Claude wrote, per its own report
  • 8x Increase in code shipped per Anthropic engineer per quarter versus 2021-2025
  • 3 Models in OpenAI’s GPT-5.6 preview: Sol, Terra, and Luna
  • 20+ Leo Platform npm packages poisoned in the June 24 worm attack
  • 34,300 GitHub stars Continue had when Cursor acquired and shuttered it
  • $1.5 Billion Baseten’s Series F, at valuations up to $13 billion
  • 82.4 SWE-Bench Verified score for the open-source Ornith-1.0 397B model
  • 85.6% CyberGym score for OpenAI’s fully released GPT-5.5-Cyber
  • June 28 CISA deadline to patch the Cisco and PTC zero-days

Quick Hits

  • Gemini Interactions API GA - June 22. Google makes it the primary interface for Gemini models and agents.
  • Anthropic When AI Builds Itself - June 22. Report says Claude wrote 80%+ of merged May code, with 8x throughput per engineer.
  • OpenAI Daybreak Expansion - June 22. Patch the Planet, an updated Codex Security plugin, and the full GPT-5.5-Cyber release.
  • Cursor Acquires Continue - June 22. Acqui-hire shuts down the 34K-star open-source coding assistant.
  • Baseten Series F - June 22. $1.5 billion raised at valuations up to $13 billion for AI inference.
  • Trigger.dev and Snowflake Outages - June 22. AWS capacity and load issues knock both offline for hours.
  • Nx Polygraph - June 23. Synthetic monorepo links repos so agents can work across them.
  • GitKraken Code Flow - June 23. Kepler agent IDE, Desktop 12 agent mode, and GitLens 18 ship together.
  • Vite 8.1 - June 23. Rolldown-powered bundler hits 41.6M weekly downloads.
  • Envoy AI Gateway 1.0 - June 23. First GA of the CNCF AI gateway, backed by Bloomberg, Nutanix, and Tetrate.
  • Security Profiles Operator 1.0 - June 23. Kubernetes seccomp and SELinux operator graduates to a stable v1 API.
  • Superhuman Buys GPTZero - June 23. AI detection startup with 19M users joins Superhuman.
  • MoEngage Buys Aampe - June 23. All-cash deal bets on per-customer AI marketing agents.
  • Leo Platform npm Attack - June 24. A Shai-Hulud worm variant poisons 20+ packages via a stolen maintainer account.
  • Walmart Tech Layoffs - June 24. 306 Sunnyvale product and engineering roles cut, effective August 21.
  • Ornith-1.0 Released - June 25. DeepReinforce open-sources MIT-licensed coding models from 9B to 397B.
  • Cisco and PTC Zero-Days on KEV - June 25. CISA sets a June 28 deadline for CVE-2026-20230 and CVE-2026-12569.
  • Amperity Layoffs - June 25. Seattle customer data startup cuts staff in an AI-driven reshape.
  • Next.js 16.3 Preview - June 25 and 26. Instant Navigations plus AGENTS.md docs, Skills, and an agent browser.
  • GPT-5.6 Sol Preview - June 26. OpenAI starts a government-gated limited preview of Sol, Terra, and Luna.

The theme this week was control, in two senses. On one side, governments are now standing between frontier models and the developers who use them, with OpenAI’s gated GPT-5.6 preview following the same pattern that pulled Anthropic’s Fable 5 offline a few weeks back. On the other, the tools themselves keep reshaping how code gets made, from Anthropic admitting Claude writes most of its codebase to Nx and GitKraken building for a world where you orchestrate a dozen agents at once. Then the supply chain showed up on schedule, with a second worm in a month chewing through an npm namespace and CISA setting another tight patch clock. More capability, more oversight, and the same old reminder that trust does not scale as fast as code. Next week the wait for Gemini 3.5 Pro continues, and GPT-5.6’s broader rollout should start coming into view. See you then.