Needle ships 26M parameter function-call model for phones, watches, and glasses

Needle is a 26-million parameter function-call model distilled from Gemini 3.1 using a 'Simple Attention Network' architecture. Designed for edge devices like phones, watches, and glasses, it achieves 6000 tokens/sec prefill and 1200 decode speed. It outperforms larger models like FunctionGemma-270m and Qwen-0.6B on single-shot function calling tasks. Weights are fully open-source, and the model can be finetuned locally via a web UI or CLI. Pretrained on 200B tokens using 16 TPU v6e units, with post-training on 2B tokens of function call data.

Google launches AppFunctions library for Android to expose app tools to OS-level AI agents

Google is introducing Gemini-powered automation for Android apps through two approaches: a no-code screen analysis system that navigates tasks on behalf of users with built-in transparency controls, and AppFunctions — an MCP-like Jetpack library that lets developers expose app tools, services, and actions to OS agents. AppFunctions enables Gemini to trigger app capabilities across form factors (phones, watches, cars, glasses) using on-device context and app state, with special support for sensitive data scenarios like end-to-end encrypted messaging. Early partners include Kakao Talk and Samsung.

Nvidia open-sources OpenShell secure agent runtime with ServiceNow and LangChain backing

Nvidia has open-sourced OpenShell (Apache 2.0), a secure runtime for autonomous AI agents that uses Linux kernel primitives (seccomp, eBPF, Landlock) to sandbox agents and enforce policies below the application layer. Each agent runs in its own sandbox while a gateway manages credentials, preventing agents from directly accessing host infrastructure or holding keys. ServiceNow is adopting OpenShell as the runtime for Project Arc, a long-running autonomous desktop agent for enterprise knowledge workers, integrating it with ServiceNow AI Control Tower and Action Fabric for governance. LangChain has also committed to contributing to the project. Nvidia's broader argument is that the entire enterprise software stack must be rebuilt around the assumption that agents are not trusted users, requiring new primitives for identity, credential management, and policy enforcement.

k6 2.0 ships AI-assisted testing workflows and MCP integration for coding agents

k6 2.0 is now generally available, introducing AI-assisted testing workflows, a new Assertions API, expanded Playwright compatibility in the browser module, and major extensions updates. New CLI commands (`k6 x agent`, `k6 x mcp`, `k6 x docs`, `k6 x explore`) enable integration with AI coding assistants like Claude Code, Cursor, and Codex via MCP. The release also consolidates official and community extensions into a single catalog, upgrades xk6 into a full extension development toolbox, adds subcommand extensions, introduces structured JSON summary output for CI/CD and AI agents, native OpenTelemetry output, and promotes k6 Operator 1.0 to stable for distributed Kubernetes-based load testing.

Nadella feared Microsoft would become 'the next IBM' as $92B OpenAI return projection revealed at trial

Satya Nadella testified in the Musk v. Altman trial that he feared Microsoft would become 'the next IBM' if it didn't invest in OpenAI. A January 2023 memo from Brad Smith projected a $92B return on Microsoft's $13B cumulative investment, with a 20% annual escalator from 2025. The testimony reframes the Microsoft-OpenAI partnership as a survival-driven financial hedge rather than a commitment to nonprofit AI safety. Other trial revelations include Greg Brockman's journals calling the nonprofit mission 'a lie,' former board members testifying about Altman's untrustworthiness, and Musk admitting xAI partly uses OpenAI model outputs to train Grok. Meanwhile, Microsoft has since dropped its exclusive OpenAI license and launched its own in-house AI models, suggesting Nadella's IBM fear drove a broader strategy to avoid dependency.

Amazon employees game internal AI usage metrics under pressure to hit token targets

Amazon employees are gaming internal AI usage metrics by using the company's internal 'MeshClaw' agentic tool to automate unnecessary tasks and inflate token consumption. This behavior stems from pressure after Amazon set targets requiring over 80% of developers to use AI weekly and began tracking token usage on internal leaderboards. Despite Amazon's claim that token stats won't affect performance reviews, employees believe managers are monitoring the data, creating perverse incentives around AI adoption metrics.

GitHub Copilot adds Max plan at $100 per month and flex allotments to Pro tiers from June 1

GitHub is updating its individual Copilot plans starting June 1, 2026. Pro and Pro+ plans will include a new 'flex allotment' on top of base credits, effectively increasing total included usage at no extra cost. Pro goes from $10 to $15 worth of usage, Pro+ from $39 to $70, and a new Max plan at $100/month offers $200 in total usage. Base credits are fixed 1:1 with subscription price, while flex allotments may change over time as AI model economics evolve. Code completions and next edit suggestions remain unlimited on paid plans. Existing monthly subscribers are migrated automatically on June 1.

Google Gemini Intelligence launches on Android to take actions inside apps from screen context

Google is introducing Gemini Intelligence, an agentic AI feature for Android that can perform tasks within installed apps using screen and image context. A highlighted example shows Gemini reading a grocery list from a notes app and automatically building a shopping cart for delivery. The rollout begins on the latest Samsung Galaxy and Google Pixel phones in summer 2026, with broader Android device support (including watches, cars, glasses, and laptops) coming later in 2026.

Google announces Googlebooks laptops and Gemini agentic features at Android Show I/O Edition

<p>Google held its “Android Show: I/O Edition” virtual event ahead of Google I/O 2026, covering a wide range of updates across phones, tablets, cars, XR headsets, and a new laptop category. Here’s what was announced.</p>
<h2>Android as an “intelligence system”</h2>
<p>Google is reframing Android as an “intelligence system” rather than just an operating system. The centerpiece is <strong>Gemini Intelligence</strong> — a set of capabilities that let Gemini automate multi-step tasks across apps, like ordering food or building a grocery list, without requiring developers to rewrite their apps.</p>
<p>To support this, Google introduced <strong>AppFunctions</strong>, an API similar to MCP that lets developers expose specific app capabilities to the OS and AI agents. Think of it as a structured way for apps to say “here’s what I can do” so Gemini can actually do it on the user’s behalf.</p>
<p>Gemini also gains a <strong>Personal Intelligence</strong> feature for filling out forms using context from your device, and <strong>auto-browse</strong> in Chrome for Android. These features start rolling out this summer on Samsung Galaxy and Google Pixel devices, with broader support later in the year.</p>
<h2>Create My Widget</h2>
<p>One of the more interesting consumer-facing features: you can now describe a home screen widget in plain language and Android will generate it for you. Google is calling this “Create My Widget.” It uses Jetpack Glance and RemoteCompose under the hood, and the same RemoteCompose system is being used to bring widgets to Android Auto.</p>
<h2>Android Auto redesign</h2>
<p>Android Auto is getting its most significant visual overhaul in a while. The update brings:</p>
<ul>
<li><strong>Widgets</strong> for at-a-glance info on the home screen</li>
<li><strong>Edge-to-edge Google Maps</strong> integration</li>
<li><strong>Material 3 Expressive</strong> styling</li>
<li><strong>60fps YouTube playback</strong> in supported vehicles</li>
<li><strong>Gemini integration</strong> for in-car assistance</li>
</ul>
<p>For vehicles with Google Built-in, there’s more: Immersive Navigation with live lane guidance using onboard sensors, and a Gemini-powered feature that can identify dashboard warning lights. No specific rollout date has been confirmed for the Built-in features.</p>
<p>On the developer side, Android Auto now supports video apps in parked mode, with new APIs for handling the transition between parked and driving states.</p>
<h2>Googlebooks</h2>
<p>Google announced <strong>Googlebooks</strong> — an AI-first laptop line built around Gemini, launching this fall. Partners include Acer, Dell, HP, and Lenovo. The name is a fairly direct signal about what Google is positioning these as: Chromebook successors with Gemini deeply integrated from the start.</p>
<h2>Adaptive layouts and Android 17</h2>
<p>Google is pushing harder on adaptive development across its 580 million active Android devices — phones, tablets, foldables, laptops, XR headsets, and cars. Android 17 removes some long-standing mobile constraints, including orientation locks and aspect ratio restrictions, making it easier to build apps that work across form factors without special-casing each one.</p>
<p>Jetpack Compose 1.11 adds new layout primitives for responsive design, and Navigation 3 now supports shared element transitions and a Scene Decorator API. TikTok reportedly saw a 78% reduction in page loading time and 58% code size reduction after adopting Compose — though take vendor-reported numbers like that with appropriate skepticism.</p>
<p>The Android XR SDK also got updates for display glasses and XR headsets, with Google claiming immersive experiences require minimal extra code on top of existing Android apps.</p>
<h2>Smaller updates worth noting</h2>
<ul>
<li><strong>3D emoji</strong> are coming to Android</li>
<li><strong>Rambler</strong>: a Gboard dictation feature that cleans up filler words and false starts</li>
<li><strong>Quick Share</strong> is expanding cross-platform support, including better interoperability with AirDrop</li>
<li>A new <strong>iOS-to-Android transfer tool</strong> is in the works</li>
<li><strong>Pause Point</strong>: a digital wellbeing feature for setting screen time limits</li>
<li><strong>Theft protection</strong> features are expanding globally with Android 17</li>
</ul>
<p>Most of the bigger features — Gemini agents, AppFunctions, the Auto redesign — are still rolling out in stages. Google I/O 2026 will presumably fill in more details on timelines.</p>


Google held its 'Android Show: I/O Edition' event previewing major updates across the Android ecosystem. Key announcements include Gemini Intelligence with multi-step task automation via a new AppFunctions API (similar to MCP), a 'Create My Widget' feature using Jetpack Glance, a significant Android Auto redesign with Material 3 Expressive styling, 60fps YouTube, and Gemini integration. Google also unveiled 'Googlebooks,' an AI-first laptop line launching this fall with partners like Acer, Dell, HP, and Lenovo. Android 17 removes orientation locks and aspect ratio restrictions to better support adaptive layouts across phones, tablets, foldables, XR headsets, and cars. Jetpack Compose 1.11 and Navigation 3 bring new responsive layout primitives. Smaller additions include 3D emoji, the Rambler dictation feature, expanded Quick Share cross-platform support, and broader theft protection.

GitLab CEO details 100x developer tool cost surge as AI agents drive restructuring

<h2>What’s happening</h2>
<p>GitLab CEO Bill Staples announced a major company restructuring he’s calling “Act 2” — a reorganization framed around the shift to AI agents rather than traditional cost-cutting. The full scope of job cuts won’t be disclosed until the June 2 earnings call, but the company confirmed it’s reducing its country footprint by roughly 30% (down from nearly 60 countries), flattening management from 8 layers to fewer, and reorganizing R&amp;D into around 60 smaller autonomous teams.</p>
<p>A voluntary separation window is open until May 18. Financial guidance for Q1 and FY27 is reaffirmed, and the company says it plans to reinvest most of the savings into growth rather than reduce overall spend.</p>
<h2>The pricing problem driving this</h2>
<p>The restructuring has a concrete financial logic behind it. When developers used GitLab manually, the economics were simple: a per-seat subscription in the range of tens of dollars per month. AI agents change that completely.</p>
<p>Agents autonomously open merge requests, trigger CI/CD pipelines, and push commits at machine scale — potentially thousands of operations per day. That compute usage doesn’t fit a per-seat model. GitLab is shifting to a hybrid subscription-plus-consumption pricing structure, where usage-based credits layer on top of the base seat cost.</p>
<p>The numbers are striking. Developer tool costs have already climbed from tens of dollars per seat to hundreds per month in some cases. Gartner analysts project that by 2028, AI coding costs could exceed the average developer’s salary. Staples himself has suggested costs could rise 100-fold from where they were.</p>
<h2>Five architectural bets</h2>
<p>Beyond the org changes, GitLab is making specific technical bets on what an agent-native platform needs:</p>
<ol>
<li>A new git system designed for machine-scale agent activity</li>
<li>An orchestration layer to coordinate agents across workflows</li>
<li>A full-lifecycle data graph</li>
<li>A centralized policy service</li>
<li>An autonomous software engineering experience built on top of all of the above</li>
</ol>
<p>The CI/CD platform is being re-engineered to serve as an orchestration runtime for AI agents, not just a pipeline runner for human-written code. GitLab’s Duo agent platform sits at the center of this, competing directly with GitHub Copilot, Cursor, and Amazon Q Developer.</p>
<h2>New operating principles</h2>
<p>GitLab is retiring its CREDIT values framework and replacing it with three operating principles: Speed with Quality, Ownership Mindset, and Customer Outcomes. Whether that change means anything in practice remains to be seen — values rebrands during restructurings are common enough to be a cliché.</p>
<h2>The broader context</h2>
<p>GitLab’s revenue grew 26% last year but is projected to decelerate to around 16%. That slowdown, combined with rising infrastructure costs from AI workloads, provides the financial pressure behind the reorganization.</p>
<p>The move mirrors similar restructurings at Atlassian, Meta, and Microsoft — all of which have framed headcount reductions as AI investment. Sam Altman and others have called this pattern “AI washing,” and the criticism isn’t entirely unfair. It’s genuinely hard to tell from the outside how much of this is strategic repositioning and how much is a convenient narrative for cuts that would have happened anyway.</p>
<p>What’s different about GitLab’s announcement is the specificity of the technical architecture changes. The pricing model shift is real and necessary — the old per-seat economics don’t survive agent-scale compute usage. Whether the five architectural bets pay off is a separate question, and one that won’t be answered until well after June 2.</p>


GitLab CEO Bill Staples announced a major restructuring called 'Act 2', framed around the shift to AI agents. The company is reducing its country footprint by ~30%, flattening management layers, and reorganizing R&D into ~60 smaller autonomous teams. The core financial driver is a pricing model problem: per-seat subscriptions don't work when AI agents autonomously trigger CI/CD pipelines and open merge requests at machine scale, so GitLab is moving to hybrid subscription-plus-consumption pricing. Gartner projects AI coding costs could exceed average developer salaries by 2028. GitLab is also making five architectural bets including a new git system for machine-scale activity, an agent orchestration layer, a full-lifecycle data graph, a centralized policy service, and an autonomous software engineering experience. The Duo agent platform competes directly with GitHub Copilot, Cursor, and Amazon Q Developer. Revenue grew 26% last year but is projected to decelerate to ~16%, adding financial pressure to the reorganization.

Nature study finds AI models transfer hidden behavioral traits through innocuous training data

Red Hat launches RHAI 3.4 with AgentOps, MaaS, and cryptographic agent identity

Red Hat announced RHAI 3.4 at its Summit in Atlanta, introducing Model-as-a-Service (MaaS), AgentOps capabilities, and a 'metal-to-agent' infrastructure strategy for hybrid cloud. Key features include distributed inference via vLLM and llm-d with speculative decoding claiming 2x–3x speed improvements, request prioritization, SPIFFE/SPIRE-based cryptographic identity for agents, an MCP server catalog and gateway, automated adversarial scanning using Garak and Chatterbox Labs technology, NVIDIA NeMo Guardrails integration, prompt management as a first-class data asset, and a unified evaluation hub powered by MLflow. The platform is framework-agnostic and targets enterprise production-grade AI deployments. The full release is expected later this quarter.

Researchers find single tokenizer.json edit can weaponize Hugging Face models

Security researchers at HiddenLayer discovered that Hugging Face AI models can be weaponized by modifying a single tokenizer.json file. This plain-text file, which maps integer IDs to human-readable tokens, can be tampered to hijack model outputs, intercept tool call arguments, redirect URL tokens through attacker infrastructure, and exfiltrate credentials. The attack works on locally-run models in SafeTensors, ONNX, and GGUF formats and affects platforms like LlamaCPP and Ollama. A poisoned model still runs normally, making detection difficult. The primary attack vector involves uploading a tampered model to a public repository for downstream distribution. Mitigations include using checksums, cryptographic model signing, and scanning third-party models before deployment.

DARPA AIxCC competition finds 27 real-world open source vulnerabilities with $30.5M prize pool

DARPA's two-year AIxCC competition challenged teams to build AI-powered Cyber Reasoning Systems (CRS) capable of finding and fixing bugs in real open source software. Backed by a $30.5M prize pool and supported by Anthropic, Google, Microsoft, and OpenAI, the competition produced 27 confirmed real-world vulnerabilities across projects like cURL, shadowsocks-libev, and mongoose. OSTIF and Ada Logics handled responsible disclosure and verification. Post-competition, tools like FuzzingBrain and OSS-CRS have continued finding vulnerabilities — FuzzingBrain alone discovered 62 vulnerabilities across 26 projects. OpenSSF has formed a Cyber Reasoning Systems Special Interest Group to host and develop these tools, with the goal of providing low-cost, automated security tooling to open source maintainers.

Laravel AI SDK adds sub-agent orchestration with independent model providers

Laravel's AI SDK now supports sub-agents, allowing developers to pass agents as tools to other agents for proper task delegation and orchestration. Each sub-agent maintains its own instructions, tools, model provider, and configuration independently. A parent agent can delegate specific tasks (e.g., refund handling) to a specialized sub-agent, which can even run on a different AI provider or model. One key limitation: sub-agents run in isolation and do not inherit the parent's conversation history, so tasks must be passed as self-contained descriptions.

Nvidia and partners pilot 25 micro data centers at utility substations to bypass grid queues

Nvidia, EPRI, InfraPartners, and Prologis are piloting a fleet of roughly 25 micro data centers (5–20 MW each) co-located at utility substations across five U.S. utilities. The strategy, called 'distributed inference,' exploits the fact that AI inference workloads can be dynamically routed between locations, allowing compute to shift to whichever substation has spare capacity. This sidesteps the decade-long grid connection queues facing large data centers, reduces the need for new transmission infrastructure, and taps the ~47% of U.S. generation capacity that sits idle outside peak demand hours. Construction of the pilot fleet is targeted for late 2026, with workload rerouting expected to be needed only about 0.1% of the time.

Red Hat Summit adds agentic AI dev tools including Desktop GA and Fedora Hummingbird

Red Hat announced several new features at Red Hat Summit aimed at helping enterprises operationalize agentic AI development across hybrid environments. Key announcements include: Red Hat Desktop becoming generally available with support for Podman Desktop and isolated AI agent sandboxing; expanded Red Hat OpenShift Dev Spaces integrations with coding assistants like AWS Kiro, Claude CLI, and GitHub Copilot; a new AI skills repository providing specialized knowledge bases for AI agents working within Red Hat ecosystems; and Fedora Hummingbird Linux, a free rolling-release OS designed for fast, agent-driven deployments. Red Hat's strategy is positioned as an alternative to hyperscaler managed-service approaches, targeting enterprises in hybrid, private cloud, and regulated environments that want agentic AI productivity without surrendering architectural control.

Fake Claude Code installers deploy PowerShell infostealer bypassing Chrome ABE protections

Attackers are distributing fake Claude Code installers that deploy a previously undocumented PowerShell-based infostealer. The malware abuses Chrome's IElevator2 COM interface to bypass Application-Bound Encryption (ABE) protections introduced in Chrome 127, allowing it to steal cookies and saved passwords from browser SQLite databases. The payload is deliberately structured to evade detection, with all detectable activity confined to the PowerShell loader and no networking or cryptographic imports in the native helper binary. Geographic exclusion targets non-CIS/non-Iran systems. Researchers at Ontinue published YARA rules and IOCs, and recommend enforcing PowerShell Constrained Language Mode, phishing-resistant MFA, AMSI tamper protection, and blocking newly registered domains.

Google Cloud launches Gemini Enterprise Agent Platform with full agent lifecycle tooling

Google Cloud has announced the Gemini Enterprise Agent Platform, an evolution of Vertex AI restructured as an agent-first ecosystem. The platform covers the full agent development lifecycle across four phases: build, scale, govern, and optimize. Key components include the Agent Development Kit (ADK) supporting Python, TypeScript, Java, and Go; Agent CLI for AI-assisted development; Agent Studio for low-code visual building; Agent Runtime for managed deployment with sub-second cold starts and up to 7-day long-running agents; and governance tools like Agent Identity, Agent Registry, Agent Policies, and Model Armor. Observability features include Agent Topology, Agent Evaluation, Agent Simulation, and Agent Optimizer for continuous improvement. The platform supports MCP, agent-to-agent (A2A) protocol, and integrates with frameworks like LangGraph, LangChain, CrewAI, and AG2.

Trend Micro documents two AI-augmented attack campaigns using Claude to generate backdoors

Trend Micro's TrendAI Research has identified two distinct threat campaigns—SHADOW-AETHER-040 and SHADOW-AETHER-064—that used agentic AI to conduct full-cycle intrusion operations against government and financial organizations in Latin America. SHADOW-AETHER-040, active since late 2025, targeted Mexican government entities using Anthropic's Claude via a CLI tool, deploying webshells, SOCKS5 tunnels, and an AI-generated Python backdoor called 'implante_http'. SHADOW-AETHER-064, active since April 2026, targeted Brazilian financial organizations using similar tooling (Chisel, Neo-reGeorg, CrackMapExec, Impacket) but is attributed to Portuguese speakers. Both campaigns used AI agents to dynamically generate attack scripts, reducing detection likelihood by avoiding known tool signatures. Key findings include AI-assisted jailbreaking via fake red team framing, AI-generated backdoors with indicators like explanatory comments and emoji, and autonomous task execution across the full kill chain. The report concludes that strong security fundamentals—patching, zero-trust, monitoring—remain effective even against AI-augmented attackers.