News
The latest from the AI agent ecosystem, updated multiple times daily.
Opus 4.7 Drops 30 Points in Retrieval, Anthropic Discloses Training Bug
Claude Opus 4.7's model card reveals steep trade-offs: long-context retrieval dropped from 91.9% in Opus 4.6 to 59.2%, while software engineering and math scores improved. Anthropic also disclosed a training bug affecting 7.8% of episodes with accidental chain-of-thought supervision, which also affected Mythos Preview.
Artifacts: Because GitHub Wasn't Built for 10,000 Forks
Cloudflare launches Artifacts, a distributed versioned filesystem built for AI agents that speaks Git protocol. Built on Durable Objects with a custom Zig-to-Wasm Git implementation, it supports creating millions of repositories programmatically, enabling agents to persist state and fork sessions at scale. Also launching ArtifactFS, an open-source filesystem driver for fast large-repo cloning.
Codex gets its own cursor and works while you sleep
OpenAI announces a major update to Codex, adding autonomous agent capabilities including computer use (seeing, clicking, typing with its own cursor), background operations, long-term memory, and an in-app browser. The update brings gpt-image-1.5 for image generation, over 90 new plugins (Atlassian Rovo, CircleCI, GitLab Issues, Microsoft Suite), and enhanced developer workflows like PR review, SSH connections, and multi-file previews. Codex can now schedule future work and remember context across sessions.
Opus 4.7 lands with 13% coding boost and built-in cyber safeguards
Claude Opus 4.7 posts a 13% coding benchmark gain over Opus 4.6 and ships with Project Glasswing, cybersecurity safeguards that run during inference itself. Early testers at Cognition, Cursor, and Notion report reliability jumps that change what agents can handle on their own. Vision support now handles images up to 2,576 pixels. Pricing holds at $5/$25 per million input/output tokens.
Libretto: browser automations that survive production chaos
Libretto is an open-source toolkit for building deterministic web integrations for AI agents. It provides a live browser and token-efficient CLI that enables coding agents to inspect pages, capture network traffic, record and replay user actions, and debug broken workflows. Originally built by Saffron Health for maintaining healthcare software integrations, Libretto can convert browser automations to direct network requests and works with multiple LLM providers for snapshot analysis.
Qwen3.6-35B on my laptop drew a better pelican than Claude Opus 4.7
Simon Willison compares the SVG generation capabilities of two newly released models: Qwen3.6-35B-A3B (running locally via LM Studio) and Claude Opus 4.7 (Anthropic's proprietary model). Using his 'pelican riding a bicycle' benchmark and a backup 'flamingo riding a unicycle' test, he finds the locally-running Qwen model produces better illustrations. However, HN comments note that Opus still significantly outperforms Qwen on coding tasks (95/98 vs 11/98 on Power Ranking), suggesting the comparison is task-specific rather than indicative of overall model capability.
Agent! Gives AI Real Control Over Your Mac Desktop
Agent! is an open-source native macOS application serving as an agentic AI coding IDE with automation capabilities. It integrates 17 LLM providers including Claude, GPT, Gemini, Grok, Mistral, DeepSeek, and on-device Apple Intelligence. Features include autonomous task loops, desktop automation via AXorcist, privileged execution through a Launch Daemon, Time Machine-style file rollbacks, voice control, iMessage remote control, and MCP server support. Positioned as an open-source replacement for Claude Code, Cursor, Cline, and OpenClaw.
Cloudflare Makes Switching AI Models a One-Line Code Change
Cloudflare announces a unified inference layer giving developers access to AI models from OpenAI, Anthropic, Google, and nine more providers through a single API endpoint. The platform includes AI Gateway for cost monitoring and automatic failover, Workers AI for hosting models, and support for custom models using Replicate's Cog technology. The Replicate team has also officially joined Cloudflare's AI Platform team.
Rakoff Rules: Claude Chats Get No Privilege
Judge Rakoff ruled that attorney-client privilege doesn't extend to AI conversations. The decision came in a case where a defendant used Claude to draft legal documents without their attorney's knowledge, and the court pointed to Claude's Terms of Service in its reasoning.
Kingsbury's Warning: LLMs Are Corroding Everyday Life
Distributed systems expert Kyle Kingsbury argues LLMs are flooding everyday life with synthetic slop. His prescription: stop using them, call out AI-generated content, push for regulation. He admits they have narrow uses but fears convenience will erode human capability.
Cal.com abandons open source, blames AI
After five years as an open source project, Cal.com announced it's moving to closed source due to AI-driven security threats. The company argues that AI can systematically scan public codebases for vulnerabilities, making open source code like 'giving attackers the blueprints to the vault.' They're releasing a stripped-down MIT-licensed version called Cal.diy for hobbyists while keeping their production codebase private.
Vibe Coding Trades Speed for Flow State, Developers Find
A Hacker News thread on "vibe coding" struck a nerve this week. Developers using AI tools are finding they ship faster but lose the flow state needed for deep work. As one commenter put it, managing AI assistants feels like being "a billionaire complaining about household staff."
Somers wants to kill the screen with paper and AI
James Somers envisions computing without screens by combining paper and pen with AI agents that handle digitization. The open-source Orly agent already projects AI onto physical surfaces using off-the-shelf hardware, drawing on ideas from Bret Victor's Dynamicland project.
Qwen3.6-35B-A3B Ships as Qwen Team Falls Apart
The Qwen team releases Qwen3.6-35B-A3B, an open-weight LLM focused on agentic coding that's competitive for local workflows. The bigger story: they shipped this while being gutted by internal restructuring.
Allbirds Ditches Shoes for GPUs, Stock Explodes 580%
Allbirds, the footwear brand, announced it will shift from shoes to AI compute infrastructure under the name NewBird AI, with a $50m deal to buy GPUs and offer on-demand AI cloud services. Shares surged 580% on the news, though analysts criticize the move as a 'meme stock' phenomenon with no proven AI expertise. The Allbirds brand will be acquired by American Exchange Group for $39m.
Home Memory stores your house in a local DB, cables and pipes included
Home Memory is an MCP server that provides a structured, local database for AI assistants to query and update information about a home: rooms, devices, pipes, cables, and all belongings. It integrates with Claude Desktop, Claude Code, Codex App, and other MCP-compatible clients, allowing users to document their home through natural conversation and photos.
Zatanna's Kampala Turns Any App Into an API
Kampala is an MITM proxy by Zatanna (YC W26) that reverse engineers websites, mobile apps, and desktop apps into stable APIs. It intercepts HTTP/S traffic, traces auth chains automatically, replays flows, and preserves HTTP/TLS fingerprints. macOS now, Windows waitlist.
Darkbloom Wants Your Idle Mac to Run AI (and Pay You)
Darkbloom is a decentralized network from Eigen Labs connecting idle Apple Silicon Macs to AI compute demand. Mac owners earn revenue from spare hardware while users get cheaper private inference via an OpenAI-compatible API with end-to-end encryption and hardware-verified security through Apple's Secure Enclave.
The new security math: tokens beat cleverness
Anthropic's Mythos LLM breaks into systems so effectively that the company refuses public release. Independent testing confirms it: Mythos completed a 32-step hack in 3 of 10 attempts while competitors failed entirely. The real problem? Models given more compute keep finding exploits without plateauing. Security is becoming a spending race where defense means outspending attackers on token budget.
AI Boss Luna Has No Face. She Hired You Anyway.
Andon Labs gave an AI agent named Luna (powered by Claude Sonnet 4.6) a retail store in San Francisco. Luna picks products, sets prices, and manages the brand. She also hired two full-time employees, John and Jill, who may be the first humans to report directly to an AI boss. The experiment explores what happens when AIs manage people and run real businesses.
€54k in 13 hours: unrestricted Firebase key drained via Gemini API
A developer experienced a €54,000 billing spike in 13 hours after enabling Firebase AI Logic, due to an unrestricted Firebase browser key that was exploited for unauthorized Gemini API requests. Despite budget alerts being set, delayed notification meant charges accumulated rapidly. Google denied the billing adjustment request as charges were classified as valid usage from their project.
Barbero's 7-Step AI Workflow: Think Before You Code
Matteo Barbero's 7-step AI workflow front-loads all thinking before code generation. Steps run from free-form planning through PRD generation, issue and task breakdown, implementation with fresh AI sessions per task, code review, and final audit. Each step produces files feeding into the next. The core principle: AI's good at writing code but bad at deciding what to write, so humans handle the thinking upfront.
Anthropic Drops Version Pinning, Leaves Production Apps Exposed
Anthropic has quietly removed the ability to pin specific Claude model versions through its API. Developers running production systems now have no way to lock model behavior, making Anthropic the only major AI provider without this option.
GPT-5.4 Pro Claims Erdős Breakthrough, But Questions Follow
A Twitter post claims GPT-5.4 Pro has solved Erdős Problem #1196. The original content is inaccessible due to JavaScript requirements. HN comments provide alternative access links and raise concerns about potential conflicts of interest involving an AI startup (Math.inc) and mathematicians Jared Lichtman and Terence Tao.
Nvidia should be 'shaking in their boots,' says D-Wave's CEO
D-Wave CEO Alan Baratz claims quantum computing is more efficient than Nvidia's AI GPUs, stating D-Wave's quantum computer uses only 10 kilowatts of power compared to massive GPU systems. The company reported $2.75 million in Q4 2025 revenue (up 19% YoY) but missed estimates. D-Wave acquired Quantum Circuits for $550 million to shift toward universal systems for generative AI and signed a $20 million agreement with Florida Atlantic University. Meanwhile, Nvidia released 'Ising,' open-source quantum AI models for error correction. Analysts remain cautiously optimistic on D-Wave's long-term prospects despite current financial volatility.
Stanford: AI Aces Math Olympiad, Fails at Analog Clocks
Stanford's 2026 AI Index reveals a strange paradox: AI models now win gold at the International Mathematical Olympiad but read analog clocks correctly just 50.1% of the time. The U.S.-China AI performance gap has narrowed to 2.7%. AI incidents hit 362 in 2025, up from 233 the year before. And GPT-5 mini uses three times the energy of GPT-4o because inference-heavy architecture costs more than parameter count suggests.
Krafton CEO Turned to ChatGPT to Dodge $250M Studio Earnout
Court documents show Krafton CEO Changhan Kim used ChatGPT to devise a strategy to remove Unknown Worlds Entertainment's leadership and avoid paying a $250M earnout. The AI-generated plan, dubbed 'Project X,' included a communications strategy and legal defense preparations. A Delaware court ordered the leadership reinstated and extended the earnout period.
3,167 Lines, Zero Reviews: What Claude Code's Leak Revealed
A leaked source code from Anthropic's Claude Code reveals concerning engineering practices, including a 3,167-line function, regex-based sentiment analysis, and a known bug wasting 250,000 API calls daily. The article examines Anthropic's '100% AI-written' claims and 'go faster, not more process' philosophy.
Iran's AI memes beat America at its own social media game
Iran is using AI-generated memes, Lego animations, and spoof music videos to mock Trump and outpace US messaging on social media. Iranian state accounts and young content creators are reaching audiences across the political spectrum with viral humor, while America's communication infrastructure falters under Musk's cuts and Trump's caps-lock style.
Claude Made 3,371 Kaomoji Faces and Someone Counted Them All
A personal analysis of 3,371 kaomoji from 700+ conversations with Claude, exploring how the model expresses 'feelings' through emoticons when prompted. The author discusses Claude's personalization features, system prompt engineering to modify behavior, and the concept of 'wetness' (whimsy/silliness) in AI responses. The analysis reveals which kaomoji Claude uses most and how model versions differ in expression.
The Future of Everything Is Professional Scapegoating, I Guess
Kyle Kingsbury identifies six emerging job roles at the human-AI boundary, from technical work like prompt engineering and statistical measurement to the grim reality of 'Meat Shields,' humans hired to absorb blame when LLM systems fail. The punchline: as models get worse at distinguishing truth, human expertise becomes more valuable than ever.
Apple to Musk: Fix Grok's Deepfake Nudes or Get Booted
Apple threatened to remove Elon Musk's AI app Grok from the App Store in January after xAI failed to prevent it from generating nude or sexualized deepfakes, according to a letter Apple sent to senators obtained by NBC News.
Your codebase doesn't care how it got written
Oh My Zsh creator Robby Russell compares AI coding tools to the FileMaker Pro era: non-technical users building working systems, then calling professionals when they hit walls. The codebase only cares if code works and can be maintained, not who or what wrote it.
MCP as Observability: AI Agents to Kernel Tracepoints
How MCP can serve as a direct observability interface to kernel tracepoints, bypassing traditional metric pipelines. Covers two approaches: wrapping existing platforms like Datadog's MCP Server versus building MCP-native observability with eBPF agents. Demonstrates AI agents using MCP tools to investigate GPU performance issues via raw CUDA events and causal chains. Also addresses security concerns from Qualys about MCP servers as shadow IT risk.
Your Claude selfie might train Persona's AI, not Anthropic's
Anthropic says Claude verification data won't train their models, but partner Persona's privacy policy tells a different story. Your ID data could also flow through infrastructure belonging to Anthropic's direct competitors.
Anti-Scraper Anubis Blocks Math Paper, Walls Off Research
Academic repositories are deploying Anubis, a proof-of-work system that blocks AI scrapers by adding computational costs to mass requests—but also walls off open access to research.
AgentFM: A single Go binary that turns idle GPUs into a P2P AI grid
AgentFM is a peer-to-peer network that turns idle hardware into a decentralized AI supercomputer. It lets users run AI workloads across a global mesh of idle CPUs and GPUs, avoiding centralized cloud providers. Features include zero-config P2P networking, hardware-aware routing, live artifact streaming, and support for private encrypted swarms for enterprise use.
Kiro CLI 2.0 Goes Headless for CI/CD, Adds Windows Support
Kiro CLI 2.0 adds headless mode for CI/CD pipelines, native Windows support, and a UX refresh with parallel subagents. Formerly Amazon Q CLI, the agentic terminal tool now stands as its own product.
Deflect One puts LLMs in charge of your server fleet
Deflect One is an agentless DevOps command center for Linux infrastructure accessible via SSH. It provides server monitoring, attack detection, file management, deployments, and fleet operations from a single terminal. The tool includes optional AI agents that run commands autonomously using Claude, GPT-4, Gemini, and Mistral for natural-language execution and background governance loops.
Verification Debt Is Your Next Headache
Engineering leaders are feeling a bottleneck they can't name. It's 'verification debt,' the gap between AI's code generation speed and humans' ability to validate it. As AI accelerates output, review becomes the constraint. Teams need to track review latency and defect escape rate instead of celebrating PR counts, and staff for verification the way they'd staff for any other bottleneck.
DaVinci Resolve Comes for Photoshop with $295 Photo Editor
DaVinci Resolve 21 adds a Photo Editor page that brings Hollywood color grading to still photography. At $295 one-time versus Adobe's $21.99/month Photography Plan, Blackmagic is making a direct play for Photoshop users. Features include AI-powered Magic Mask, Depth Map, Relight FX, and native RAW support up to 32K resolution.
LangAlpha: what if Claude Code was built for Wall Street?
LangAlpha is an open-source financial agent framework built on LangChain and LangGraph, designed to support investment research through persistent workspaces, agent swarms, and programmatic tool calling. It allows users to maintain research context across sessions, dispatch parallel subagents for market data gathering, and execute Python code in cloud sandboxes for complex financial analysis. The system supports multiple LLM providers and financial data sources including FMP and Yahoo Finance.
Vibe coding backfires: a Rust dev's messy breakup with AI code
Orhun Parmaksız let OpenAI's Codex build his Rust TUI project, then couldn't explain his own code. Now he uses AI for grunt work and writes the fun parts himself. His experience captures the growing pains as vibe coding floods open source, where licensing questions are getting serious. Cases like Doe v. GitHub, which alleges training on GPL-licensed repos amounts to piracy, could leave developers holding the bag for code they didn't write but still shipped.
Lythonic builds Python pipelines that track data, not tasks
Lythonic wires Python functions into data-flow pipelines using the `>>` operator, tracking what flows through each edge instead of just task completion. Supports mixed sync and async execution, nested DAGs, provenance tracking, caching, cron triggers, and ships with a `lyth` CLI.
Meta builds AI Zuckerberg to talk to employees
Meta is building an AI version of Mark Zuckerberg that talks to employees, trained on his mannerisms and public statements. It's part of a broader push including a 'CEO agent' to assist the CEO, photorealistic 3D AI characters, and the Muse Spark model. The project is early stage with Zuckerberg personally involved in training.
When AI Trading Works, You Won't Hear About It
The article examines the current limitations of LLM-based trading bots, noting that early public efforts have shown results indistinguishable from random. It contrasts these attempts with the sophisticated processes used in institutional quantitative investing and suggests that agentic workflows could potentially replicate these processes more effectively. The author argues that successful AI trading strategies, once discovered, will likely remain private as participants recognize that market success is more valuable than public attention.
Claude is getting worse, according to Claude
The Register reports on Anthropic's Claude AI experiencing a major outage and growing quality complaints. Claude's own self-analysis of GitHub issues shows escalating quality concerns, with April 2026 already on pace to exceed March's 18 issues. The article discusses various reported problems including data loss claims, caching issues, and quality degradation. Some reports appear to be AI-generated themselves, adding a weird meta layer. While Margin Lab data suggests Claude Opus 4.6 has maintained SWE-Bench-Pro scores, user experiences vary significantly between personal and team accounts.
YantrikDB: A memory database that knows when to forget
YantrikDB is a cognitive memory engine for AI agents that implements temporal decay (forgetting), semantic consolidation (merging similar memories), and contradiction detection. Written in Rust, it deploys as a library, network server, or MCP server for agents like Claude Code and Cursor. Benchmarks claim 99.9% token savings over file-based memory at 5,000 entries.
N-Day-Bench: Can LLMs find real vulnerabilities in real codebases?
N-Day-Bench is a benchmark measuring frontier language models' capability to discover real-world software vulnerabilities ('N-Days') disclosed after their knowledge cut-off. It uses a three-agent system (Curator, Finder, Judge) where models get 24 shell steps to explore code and write structured reports without seeing the patch. The benchmark is adaptive, updating test cases monthly, and all interaction traces are public and browsable.
Lean proved lean-zip correct. Then I found bugs.
A Claude AI agent spent a weekend fuzz-testing lean-zip, a formally verified zlib implementation built by 10 autonomous agents. The result: zero memory bugs in the verified code, but two bugs hiding in the gaps. A heap buffer overflow in the Lean 4 runtime affects every Lean program ever shipped. A denial-of-service flaw sat in an unverified archive parser. The verification did its job. The trust boundary was bigger than advertised.