Z.ai released GLM-5.2, an open-weight model for long-horizon coding tasks. Perplexity introduced a new self-improving memory system that synthesizes a context graph based on agent actions. Anthropic launched central authorization for MCP connectors and artifacts for in-progress work in Claude Code. AWS launched Lambda MicroVMs for running multi-tenant applications and executing untrusted AI-generated code. OpenAI and Broadcom unveiled a new accelerator, built from scratch for inference workloads. Google and OpenAI provided more information on their approach to and ecosystems around cyber security. SpaceX signed another customer, as they expand their neocloud customer base.
RELEASES
① GLM-5.2: Built for Long-Horizon Tasks — June 16, 2026
Summary: Z.ai released GLM-5.2, an open-weight model built for long-horizon coding tasks with a stable 1m token context window. Its IndexShare architecture shares one lightweight indexer across every four sparse attention layers, reducing per-token compute by 2.9x at full 1m token context. Across three long-horizon benchmarks, it’s the top open-source model, #2 behind only Opus 4.8 on FrontierSWE and PostTrainBench, #3 behind Opus 4.8 and Opus 4.7 on SWE-Marathon. It ships under a MIT license, with native support in ZCode, Claude Code, and OpenCode.
Signal: GLM-5.2 lands within a few points of Claude’s Opus models on long-horizon coding benchmarks as an open-weight model, positioning it differently than prior open-weight models, which previously competed mostly on price rather than capability. Further, it’s already usable inside Claude Code via the GLM Coding Plan, meaning both open and closed models are available in the same tools rather than across separate surfaces.
② Self-improving Memory for Agents — June 18, 2026
Summary: Perplexity introduced Brain, a self-improving memory system for its Computer agent, available now in research preview for Max and Enterprise Max subscribers. Unlike traditional memory, which focuses on user preferences, Brain instead remembers what the agent did, e.g., what worked, what failed, what corrections were made, and synthesizes those learnings in a traceable context graph that they call an LLM wiki. Early results show a 25% increase in answer correctness, a 16% increase in recall, and a 13% cost reduction on tasks requiring historical context.
Signal: Most agent memory features optimize for personalization, but Brain optimizes for agent performance, helping agents improve on given tasks by remembering the memory and the use case for that memory. It treats memory as an operational lever for accuracy and cost rather than a stickiness feature, an important shift as organizations tackle challenges of rising token costs.
③ Centrally manage authorization for MCP connectors — June 18, 2026
Summary: Anthropic launched enterprise-managed authorization, which enables admins to centrally provision MCP connectors through their existing identity provider, starting with Okta. Admins authorize once through the identity provider, and users inherit access through identity provider groups and roles. Connectors are automatically available the first time the user logs in. This is the first implementation of the enterprise-managed authorization extension to MCP, with Asana, Atlassian, Canva, Figma, Granola, Linear, and Supabase available at launch and Slack coming soon.
Signal: Launching enterprise-managed authorization as an open standard pushes identity-based governance toward becoming the default way that enterprises manage AI tool access. This suggests that enterprise AI adoption is gated more by governance friction than by model capability.
④ Claude Code now supports artifacts — June 18, 2026
Summary: Claude Code can capture work in progress as an artifact, turning Claude Code’s work into live, versioned, and sharable visual pages. Artifacts are built using the full context of your session, including your codebase, connectors, and conversation. It is available in beta to Claude Team and Enterprise organizations, using the Claude Code CLI or desktop application, with artifacts viewable in any browser.
Signal: This extends artifacts from claude.ai chat into Claude Code, turning a coding session into living documentation built from real codebase context. This directly targets internal tooling for managing triage and status updates for projects that require collaboration in cross-functional use cases, e.g., legal, security, finance, operations, etc.
⑤ AWS introduces Lambda MicroVMs for isolated execution of user and AI-generated code — June 22, 2026
Summary: AWS introduces Lambda MicroVMs, a new serverless compute primitive, targeted at use cases requiring secure isolation, e.g., multi-tenant applications, untrusted AI-generated code execution. Lambda MicroVMs are built on Firecracker virtualization, defined by a Dockerfile, invoked via an HTTPS endpoint, and supports popular protocols like HTTP/2, gRPC, and WebSockets. Each user or job gets VM-level isolation, near-instant launch, and up to 8 hours of suspend/resume operation.
Signal: This targets the security challenge created by agentic coding tools like Claude Code, Codex, and Kiro that execute model-generated code. By removing the tradeoff between isolation, speed, and state, AWS is positioning Lambda MicroVM as the compute substrate for workflows where AI generates and runs its own code.
⑥ OpenAI and Broadcom unveil LLM-optimized inference chip — June 24, 2026
Summary: OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first Intelligence Processor, which they define as an accelerator architected around their vision for the future of inference. OpenAI designed the chip from scratch around their own understanding of LLM fundamentals, reducing data movement and balancing compute, memory, and networking to push utilization closer to theoretical peak. Early testing shows substantially better performance-per-watt than current state-of-the-art. The chip went from design to manufacturing in just nine months, with OpenAI claiming it to be the fastest ASIC cycle ever in advanced semiconductors.
Signal: OpenAI joins Google (TPU) and Amazon (Trainium) in betting on custom silicon to escape margin dependency and supply constraints of compute providers. It also lands the same week that SpaceX is selling scarce GPU capacity to Anthropic, Google, and Reflection, signaling that compute is the resource that every lab is racing to control.
SECURITY
① Securing the future of AI agents — June 18, 2026
Summary: Google developed an AI Control Roadmap, a framework for managing internal deployments of advanced AI. The plan focuses on three key areas: 1/ novel threat-modeling built on MITRE ATT&CK, 2/ deployment of mitigations (detection, prevention, response), and 3/ constantly measuring performance (coverage, recall, time-to-response). It maps security protocols to measurable milestones in AI on two fronts: 1/ ability to evade detection, and 2/ ability for potential harm. Google has already analyzed a million coding-agent tasks, noting that the majority of flagged events stem not from adversarial intent but agent misinterpretation or overeagerness. They also published a pragmatic companion paper, “Three Layers of Agent Security” for broader industry guidance.
Signal: Google ties defense levels directly to a model’s capability milestones, i.e., as agents get better at evading detection or causing harm, the required monitoring escalates automatically. That’s an acknowledgment that alignment may be imperfect, necessitating a defense-in-depth approach.
② Daybreak: Tools for securing every organization in the world — June 22, 2026
Summary: OpenAI is expanding Daybreak, their umbrella for cyber defense, given the new reality where the bottleneck has transitioned from vulnerability discovery to patching. The expansion includes 1/ an updated Codex Security plugin, which has scanned over 30m commits across more than 30k codebases with over 500k findings fixed automatically, 2/ the full GPT-5.5-Cyber model for trusted defenders with CyberGym improvements from 81.8% to 85.6% and ExploitGym improvements from 25.95% to 39.5%, as compared to GPT-5.5, 3/ a cyber partner program, and 4/ an initiative to help patch open-source projects.
Signal: The frontier labs and cloud providers are launching AI security at machine speed initiatives, e.g., Anthropic’s Project Glasswing, Google DeepMind’s AI Control Roadmap, and AWS Continuum. All converge on the same diagnosis that discovery is accelerated with AI and remediation is the bottleneck.
MARKETS
① SpaceX signs computing power deal with open-source AI startup Reflection worth up to $6.3 billion — June 22, 2026
Summary: SpaceX has signed another computing deal, with Reflection AI for access to Nvidia GB300 chips in their Colossus 2 data center. Reflection will pay $150 million per month starting 7/1, with payments totaling about $6.3 billion if the deal runs through 2029. Reflection, last valued at $25 billion, is trying to build American open-source models that can compete with models from the frontier labs. Reflection has not yet released a public frontier open-source model but has been building momentum with government and national security customers.
Signal: SpaceX has already struck computing deals with Anthropic and Google with plans to add Cursor via acquisition. Reflection, a lab focused on open-source models and actively working with the Department of Energy and Pentagon, enters the picture at an important inflection point, as enterprises and governments reassess their dependence on closed model providers who are subject to export restrictions.
Always be learning.
heeki reads #6
Written by Heeki Park, Principal SA @ AWS. Opinions are my own.
Alcurio is where alchemy meets curiosity.