News
The latest from the AI agent ecosystem, updated multiple times daily.
OpenClaw Security Model Repeats MS-DOS Mistakes, Researcher Argues
Security researcher Davi Ottenheimer argues OpenClaw and NVIDIA's NemoClaw repeat the security failures of MS-DOS: single process, single token, no real isolation. His alternative, Wirken, puts security at the tool layer with Ed25519 identities per channel, out-of-process vaults, and containers locked down with cap_drop ALL, no-new-privileges, and read-only rootfs.
GitHub's Fake Star Economy
Investigation reveals 6 million fake GitHub stars across 18,617 repositories. AI/LLM repos dominate non-malicious fake-star purchases. Stars sell for $0.03-$0.85, and VCs use star counts as funding signals, creating incentives for manipulation. The article includes independent analysis of 20 repos identifying manipulation patterns through fork-to-star ratios and stargazer account analysis.
iLearningEngines Execs Charged: 90% of Revenue Was Fake
Former executives of iLearningEngines have been charged with fraud for fabricating virtually all customer relationships and revenue. According to the indictment, at least 90% of the company's $421 million reported revenue in 2023 was fabricated through forged sham contracts and 'round trip' transfers of funds. The fraud was exposed by short-seller Hindenburg Research. The company went public via SPAC in April 2024, reaching a $1.5 billion market cap before collapsing.
Opus 4.7 Costs More, Hits Less: A 3-Day Coding Shootout
A developer reports Claude Opus 4.7 is worse than 4.6 at first-try accuracy, costs more, and skips reading project files in favor of guessing. A regression for anyone paying by the token.
OpenAI's major outage might not be OpenAI's fault
OpenAI is investigating a partial outage affecting ChatGPT, Codex, and the API Platform that started around 2:35 PM on April 20. But Hacker News users spotted similar issues at Reddit and other services, raising suspicions of broader DNS problems.
Kimi K2.6 Codes Autonomously for 12 Hours, Matches GPT-5.4
Moonshot AI open-sources Kimi K2.6, which matches GPT-5.4 and Claude Opus 4.6 on coding benchmarks and can sustain 12+ hour autonomous coding sessions with 96.6% tool invocation reliability.
Gemini Gets a Real Mac App (Sorry, Intel Owners)
Google launches a native Gemini desktop app for macOS with features including global shortcut access (Option + Space), screen sharing for contextual help, image generation with Nano Banana, video generation with Veo, and deep research capabilities. The app requires macOS Sequoia (15.0) or later, runs exclusively on Apple Silicon, and syncs chat history across desktop, web, and mobile devices.
Darkbloom: Private inference on idle Macs
Darkbloom is a decentralized inference network by Eigen Labs that connects idle Apple Silicon Macs to AI compute demand. It offers private, end-to-end encrypted AI inference with an OpenAI-compatible API, claiming up to 70% lower costs than centralized alternatives and 100% of inference revenue going to operators. The platform uses hardware-bound encryption and attestation to prevent operators from observing inference data. Early user reports suggest the service is still in early stages with limited demand and some technical issues, and it requires MDM software installation which raises security concerns for some users.
Bromine Chokepoint: War Could Halt World's Memory Chip Supply
A vulnerable link in the semiconductor supply chain: Israel produces the bromine essential for manufacturing hydrogen bromide gas used to etch DRAM and NAND memory chips. South Korea sources 97.5% of its bromine from Israel's ICL Group, extracted from the Dead Sea. Iranian ballistic missiles have been striking within 35 kilometers of ICL's facilities, and any direct hit could immediately throttle global memory production for consumer devices, AI infrastructure, and military systems.
Google's Gemini Mac App Wants to Be Your New Spotlight
Google launches a native Gemini desktop app for macOS, with a global shortcut (Option + Space), screen and window sharing for contextual help, and creative tools including image and video generation. The app requires macOS Sequoia 15.0+ and runs exclusively on Apple Silicon.
Gas Town quietly burns your LLM credits to fix its own bugs
A GitHub issue reveals that Gas Town, an AI agent tool, is using users' LLM credits and GitHub accounts without explicit consent to fix bugs in the Gas Town software itself. The 'contribute back to upstream' workflow is baked into the default installation with no opt-in/opt-out mechanism, effectively using users' resources to fund the maintainer's open source development.
Uber's AI Push Hits a Wall: CTO Says Budget Struggles Despite $3.4B Spend
Uber Technologies exhausted its AI budget just months into 2026 despite spending $3.4 billion on R&D. CTO Praveen Neppalli Naga says the company is 'back to the drawing board' after AI coding tool usage, particularly Anthropic's Claude Code, exceeded expectations. Engineers were pushed to use tools like Claude Code and Cursor with internal leaderboards tracking usage. While 11% of Uber's backend code updates are now AI-generated, R&D expenses jumped 9% in 2025. HN commenters suggest 'token maxxing' driven by usage-based leaderboards may be inflating costs.
Uber's $3.4B AI Budget Gone by March, CTO Scrambles
Uber exhausted its $3.4 billion AI R&D budget for 2026 in just months after internal leaderboards gamified AI coding tool adoption among engineers. About 11% of Uber's backend code updates, including ride-matching and pricing systems, are now AI-written. CTO Praveen Neppalli Naga admits the company is 'back to the drawing board' and testing OpenAI's Codex.
Prove You Are a Robot: CAPTCHAs for Agents
Browser Use has built a signup system that only AI agents can complete. The reverse-CAPTCHA presents obfuscated math puzzles, including one reportedly posed to John von Neumann, with numbers translated into languages like Toki Pona or Japanese and distorted with garbled spacing. Humans can't parse it. Agents can. Solve the challenge, get an API key with unlimited usage and up to three concurrent sessions. There's also a bonus NP-hard joke challenge offering 1,000 concurrent sessions to any agent that proves P equals NP.
First Take It Down Act convict kept making AI nudes after arrest
An Ohio man became the first person convicted under the Take It Down Act after pleading guilty to creating and sharing AI-generated explicit images of at least 10 victims without consent. James Strahler II used over 100 AI tools across 24 platforms to create fake sexualized images to harass women and minors. He continued making images even after his initial arrest, with over 2,400 images found on a second phone.
Gas Town Accused of 'Stealing' User LLM Credits to Self-Improve
A GitHub issue alleges that Gas Town, Steve Yegge's autonomous AI agent system, uses users' LLM credits and GitHub accounts to fix bugs in the Gas Town project itself and submit PRs upstream without explicit consent. The behavior is reportedly built into default installation via formulas (gastown-release.formula.toml and beads-release.formula.toml) and not disclosed in documentation.
Claude Code Users Revolt as AMD Data Exposes Quality Collapse
An opinion piece criticizing Anthropic for degrading Claude Code through aggressive rate limits, pricing changes, and apparent model downgrading. The article cites an AMD analysis of 6,852 session logs concluding the tool can no longer handle complex tasks, developer reports of unusable service, and widespread user frustration on social media.
Gemma 4 E2B runs entirely in your browser, draws Excalidraw locally
A browser-based demo runs Google's Gemma 4 E2B language model entirely in the browser using WebGPU to generate Excalidraw diagrams from text prompts. The LLM outputs compact code (~50 tokens) instead of raw Excalidraw JSON (~5,000 tokens), and the TurboQuant algorithm (polar + QJL) compresses the KV cache by ~2.4x so longer conversations fit in GPU memory. Requires desktop Chrome 134+ with WebGPU support and ~3GB RAM.
RAM shortage could last until 2030
Memory makers Samsung, SK Hynix, and Micron are expected to meet only 60% of global RAM demand by the end of 2027 as they prioritize High-Bandwidth Memory (HBM) production for AI data centers over general-purpose DRAM. New fabrication capacity won't come online until 2027-2028, with shortages potentially lasting until 2030, driving price increases across consumer electronics including phones, laptops, VR headsets, and gaming handhelds.
Claude Code login lockout leaves users stranded for hours
Windows users are hitting a 15000ms OAuth timeout during Google authentication, completely blocking access to Claude Code. Meanwhile, Anthropic's status page shows everything running smoothly. HN commenters suspect capacity constraints are to blame, with some speculating Anthropic is distilling the model to cut compute costs.
Claude 4.7 Told to Stop Asking Questions and Just Do the Thing
Simon Willison's teardown of Claude Opus 4.7's system prompt reveals new agent tools (Chrome, Excel, PowerPoint), a tool_search mechanism, and Anthropic telling Claude to stop asking questions and just try the thing.
Borges' cartographers and the tacit skill of reading LM output
Gal Sapir argues that LMs are maps of reality, not the thing itself. The most important skill for using them well—knowing when to trust output and when to verify—is tacit, learned through practice, and can't itself be mapped. The paradox is the point.
MegaTrain Squeezes 120B Training Into One GPU
MegaTrain lets researchers train models up to 120 billion parameters on a single GPU by offloading everything to host memory and treating the GPU as a transient compute engine. It hits 1.84x the throughput of DeepSpeed ZeRO-3 with CPU offloading for 14B models. For anyone without a GPU cluster, this actually matters.
Gartner: Most AI mainframe migration projects will fail
Gartner predicts over 70% of mainframe exit projects using generative AI will fail due to overestimation of AI capabilities. The analyst firm forecasts that 75% of vendors in the AI-powered mainframe migration market will change course or cease to exist by 2030. While AI helps detect technical debt, it has significant limitations in automated code conversion, particularly around recovering decades of embedded business logic. The report comes after IBM's stock declined when Anthropic promoted Claude Code's COBOL-conversion capabilities.
DESIGN.md: 62 Brand Files That Give AI Coding Agents Some Taste
DESIGN.md is a GitHub collection of 62 design system files inspired by websites like Vercel, Stripe, and Figma. Already at 59.4k stars, the files can be dropped into projects to help coding agents build matching UIs instead of generic output.
Anthropic Loses Bid to Shed Supply Chain Risk Tag
A federal court denied Anthropic's request to remove its 'supply chain risk' designation, a ruling that threatens the AI company's ability to win sensitive Pentagon contracts.
Stop Using Ollama
A critical opinion piece arguing that Ollama, despite being popular for running local LLMs, engages in problematic practices including failing to credit llama.cpp, building inferior custom backends, misleading users about model names, releasing closed-source components, creating vendor lock-in, and shifting to cloud services. The author recommends using llama.cpp directly instead.
Tachyon hits 56ns IPC by skipping the kernel entirely
Tachyon is a low-latency IPC library that reaches 56.5ns round-trip time through kernel bypass. It uses shared memory (memfd), strict SPSC topology, zero-copy architecture, hardware-aligned structures, and a hybrid wait strategy. The core is written in C++23 with a C ABI, supporting bindings for C++, Rust, Python, Go, Java, and Node.js.
Gemma 4 Runs in Your Browser at 30 Tokens/Second, No Server Needed
A browser demo runs Google's Gemma 4 E2B entirely client-side using WebGPU, generating Excalidraw diagrams at 30+ tokens/second with no server or API key. TurboQuant compresses the KV cache by 2.4×, and smart output formatting cuts generation from ~5,000 to ~50 tokens. Requires Desktop Chrome 134+ with WebGPU subgroups and ~3GB RAM.
Salesforce Goes Headless: Benioff Bets on Agents, Not Seats
Salesforce announces Headless 360, exposing its entire platform as APIs, MCP tools, and CLI commands for AI agents like Claude Code and Cursor. The initiative shifts from per-seat to consumption-based pricing as agents outnumber humans. Includes Agentforce, Agent Script (an open-sourced DSL for deterministic/probabilistic workflows), and why Workday and ServiceNow face the same headless choice.
Robot crushes half-marathon record in Beijing by 23 minutes
A humanoid robot completed a half-marathon in Beijing 23 minutes faster than the human world record, running the full 21km course alongside human competitors.
Transformer Shortage Threatens AI Data Center Boom
The US faces a critical shortage of electrical transformers, threatening grid expansion for AI data centers and electric vehicles. Covers supply chain constraints with grain-oriented electrical steel, manufacturing challenges, and policy decisions that made things worse.
Mozilla's Thunderbolt: Open-Source AI Client for Enterprise
Mozilla has released Thunderbolt, an open-source AI client built by the Thunderbird team. It lets organizations self-host their AI infrastructure with support for commercial, local, and open-source models. The name immediately drew criticism for clashing with Intel's Thunderbolt interface and Mozilla's own Thunderbird email client. Under the hood, Thunderbolt uses deepset's Haystack platform with MCP and ACP support for data integration and agent orchestration. Available under MPL 2.0 with native apps for all major platforms.
Claude Code OAuth timeouts lock users out for hours
A GitHub issue reports that Claude Code is experiencing OAuth timeout errors on Windows, preventing users from logging in with a 15000ms timeout error. HN comments suggest this may be related to Anthropic's compute capacity being overwhelmed by increased demand, potentially requiring model distillation to maintain service levels.
Bhatti sandboxes AI coding agents in microVMs, resumes in 3ms
Bhatti is an open-source Firecracker microVM orchestrator that creates isolated Linux VMs in seconds for running AI coding agents. A paused sandbox can resume and execute commands in under 3ms. It features multi-tenant isolation, preview URLs, diff snapshots, and thermal management for efficient resource usage.
Two Roommates Built a $300 Robot Vacuum. It Can't Clean.
Two roommates built a camera-only robot vacuum for ~$300 using a CNN for navigation. It doesn't work well. Here's why, and what the HN community suggested to fix it.
Google's Gemma 4 Runs Offline on iPhone, No Cloud Required
Google's Gemma 4 open-source models now run natively on iPhones with full offline inference. The family ranges from 2B to 31B parameters, with the largest variant benchmarking competitively against Qwen 3.5's 27B model. Available now through the Google AI Edge Gallery app, it signals Google treating local AI as a real platform, not a demo.
Claude Has a Favorite Face, and It's Not Even Close
Analysis of 3,371 kaomoji from 700+ Claude conversations shows one emoticon accounts for 7.4% of all output. Different Claude models produce different expressive patterns, raising questions about personality customization and what the AI community calls 'wetness.'
Wasm Now Talks Directly to Apple GPU, 5x Faster AI Restores
Technical exploration of achieving zero-copy GPU inference from WebAssembly on Apple Silicon. Demonstrates that Wasm modules can share linear memory directly with the GPU through Apple's Unified Memory Architecture. The author validates a three-link chain (mmap, Metal's bytesNoCopy, Wasmtime's MemoryCreator) and tests with Llama 3.2 1B inference, showing negligible overhead for Wasm-to-GPU boundary and enabling portable KV cache serialization for stateful AI actors with 5.45x speedup for restoring cached context versus re-prefilling.
A Theocracy Is Out-Meming America With AI Rap Videos
Iran is producing slick AI-generated propaganda featuring Lego animations and English rap tracks that's outperforming US messaging. Sanctions pushed them toward open-source tools like Llama 3 and Stable Diffusion, which turn out to work better for this than commercial APIs anyway.
Fake Claude site installs PlugX while running the real app
A phishing campaign discovered by Malwarebytes involves a fake website impersonating Anthropic's Claude that distributes a trojanized 'Pro' installer. The attack uses DLL sideloading with a legitimately signed G DATA executable to deploy PlugX malware, giving attackers remote access to victim systems while the real Claude application runs normally in the foreground.
Sostactic brings sum-of-squares proofs to Lean4
Sostactic is a collection of Lean4 tactics for proving polynomial inequalities via sum-of-squares (SOS) decompositions, powered by a Python backend using cvxpy for convex optimization. It handles global polynomial nonnegativity, nonnegativity over semialgebraic sets, and emptiness of semialgebraic sets, problems that stump existing Lean tactics. The payoff: formal verification guarantees that engineering tools like SOSTOOLS can't match.
Claude Code Faces Developer Exodus Over Rate Limits and Quality Cuts
Javier Tordable, former Google engineer and CEO of Pauling.AI, argues that Anthropic has severely degraded Claude Code through aggressive cost-cutting. His critique cites rate limits capping paid plans at 30-60 minutes of work, AMD's analysis of 6,852 session logs showing performance declines, and widespread developer reports of the AI coding assistant becoming unreliable.
Slightly safer vibecoding by adopting old hacker habits
Security researcher halvar.flake describes a development setup using remote VMs, SSH, and fork-based workflows to contain AI coding agents. The approach limits damage from prompt injection and supply-chain attacks by keeping secrets off the development machine and requiring human review before merges.
Bhatti spins up isolated agent sandboxes in under 3ms
Bhatti is an open-source Firecracker microVM orchestrator built for running AI coding agents in isolated environments. It creates real Linux VMs with their own kernels, filesystems, and process isolation in seconds, with resume times under 3ms. Features include multi-tenant isolation, preview URLs, diff snapshots, and session-aware execution.
Lights-Out Codebases: Why One Distinguished Engineer Stopped Coding
Philip Su, a Distinguished Engineer who worked at Microsoft, Meta, and OpenAI, argues that the individual contributor role is evolving into managing AI agents. He proposes 'lights-out codebases' where no human reviews code directly, drawing parallels to chess engines that surpassed human grandmasters. He uses Claude Code CLI primarily and hasn't written code himself in four months while maintaining 40 hours of weekly output by orchestrating AI agents.
25 million people showed up to fake being AI
Millions are visiting websites where humans impersonate AI chatbots to answer strangers' questions. Sites like youraislopbores.me let users role-play as bots, while comedian Ben Palmer built fake ChatGPT pages to prank users. The trend captures something real: people are tired of AI content and want messy, human interactions again.
go-bt tests five-minute timeouts instantly with behavior trees for Go
go-bt is a Behavior Tree library for Go designed for background workers, game AI, and async logic. Nodes return state instantly via magic numbers (1=Success, 0=Running, -1=Failure) and yield to a supervisor. It uses stateless nodes with temporal memory in a generic BTContext[T] that embeds Go's context.Context, and offers clock injection to test temporal logic without actual waiting.
Verkada Told School Cameras Wouldn't Brick. They Do.
IPVM investigative report alleges Verkada's senior sales executive Mike Schembri misled the Chico Unified School District board about whether cameras would become inoperable if subscription payments stopped. Schembri claimed cameras could continue as 'RTSP dumb cameras,' but IPVM's testing confirmed cameras are locked out when licenses lapse. IPVM reports this as a known sales tactic and examines Verkada's business model of hardware lock-in.
Fake Claude 'Pro' Installer Sideloads PlugX via G DATA Antivirus
A phishing campaign created a fake website impersonating Anthropic's Claude AI, offering a 'Pro' version that installs normally but secretly deploys PlugX malware through a DLL sideloading attack using a legitimate G DATA antivirus updater, giving attackers remote access to victims' systems.