News
The latest from the AI agent ecosystem, updated multiple times daily.
GPT-5.5: Mythos-Like Hacking, Open to All
XBOW's benchmarks tell a clear story: GPT-5.5 misses 10% of vulnerabilities (GPT-5 missed 40%, Opus 4.6 hit 18%). Running blind, it beats GPT-5 with full source code access. Login speeds doubled, and the model knows when to quit. It performs like Anthropic's Mythos, but you can actually use it.
Claude Desktop Quietly Installs Native Messaging Bridge
Anthropic's Claude Desktop App installs a native messaging bridge enabling browser extensions to communicate with the local app. Users must approve a Chrome permission prompt, but the desktop installer doesn't prominently explain the component, leaving some security-conscious users unsettled.
My phone replaced a brass plug
Vadim Drobinin describes building an iOS app that uses computer vision to automate shooting target scoring. The solution combines OpenCV for structural geometry and a fine-tuned YOLOv8 model (exported to CoreML) for bullet hole detection, replacing traditional physical gauges.
TSRX Compiles Cleaner UI Code for Humans and AI
TSRX is a TypeScript language extension for building declarative UIs designed for what creator Dominic Gannaway calls the 'agentic era.' It keeps structure, styling, and control flow co-located and readable while remaining fully backward compatible with TypeScript. The compiler targets React, Solid, and Ripple, handling framework-specific patterns like React hooks rules automatically to improve code readability for engineers and AI systems alike.
Anthropic Quietly Drops Claude Code from $20 Pro Plan
Anthropic has removed Claude Code from its $20/month Pro plan for new signups, with support docs now referencing only the Max Plan. The company calls it 'a small test on ~2% of new prosumer signups,' but documentation changes suggest something broader. Current Pro subscribers still have access through the web app and CLI, for now.
Firefox 150: Mythos AI Caught 271 Bugs Before Ship
Anthropic's Mythos Preview AI found 271 security vulnerabilities in Firefox 150 before release, a sharp jump from the 22 bugs caught by Anthropic's Opus 4.6 in Firefox 148. Mozilla's Bobby Holley called Mythos "every bit as capable" as the world's best security researchers, while Mozilla CTO Raffi Krikorian warned that open source maintainers still lack access to such tools.
$60B for a Code Editor? SpaceX Is After the Data
SpaceX has stated it holds an option to acquire Cursor, an AI code editor startup, for $60 billion, according to a Reuters report.
Almanac Builds Wiki Where AI Drafts, Humans Verify the Long Tail
Almanac is a wiki platform where users build knowledge bases using AI tools like Claude, ChatGPT, Cursor, and Codex. An MCP extension turns Claude Code into a research agent that drafts and submits entries. A CLI handles terminal-based contributions. Articles are attributed to contributors and opened for community edits, covering niche topics traditional encyclopedias miss.
OpenAI quietly drops Euphony, a debugging lens for Codex sessions
Euphony visualizes chat data and Codex session logs, giving developers a free way to debug AI interactions instead of guessing what went wrong.
GitHub Copilot pauses sign-ups as agentic compute costs surge
GitHub paused new sign-ups for Copilot Pro, Pro+, and Student plans as agentic workflows push compute costs past what flat-rate pricing can sustain. Opus models drop from Pro plans, billing shifts to per-token, and usage limits now show in VS Code and CLI.
CrabTrap: When your AI security guard is another AI
CrabTrap is an open-source HTTP proxy from Brex that secures AI agents in production by intercepting requests, evaluating them against policies, and allowing or blocking them in real time. It combines static rule matching with LLM judgment to make security decisions.
Some secret management belongs in your HTTP proxy
AI agents given direct access to API keys create security headaches. Some models refuse requests with visible secrets, others store keys in memory across sessions. The fix is an HTTP proxy that intercepts requests and injects authentication headers, so agents never touch the actual credentials. exe.dev's Integrations feature automates this pattern, including a GitHub App for OAuth.
MCPorter Makes MCP Servers Actually Callable
MCPorter is a TypeScript toolkit for the Model Context Protocol that discovers your existing MCP servers and generates typed clients and CLI wrappers from them. It handles OAuth, supports HTTP, SSE, and stdio transports, and lets you call MCP servers directly from TypeScript or the command line.
GPT-5.4 fixes bugs by rewriting your code. Claude doesn't.
A new research article defines 'over-editing' in AI coding models, where LLMs rewrite more code than necessary when fixing bugs. The paper introduces Token-level Levenshtein Distance and Added Cognitive Complexity to measure this behavior, and benchmarks major models including GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro Preview. Claude Opus 4.6 performs best with minimal edits, while GPT-5.4 shows the highest tendency to over-edit. Explicit prompting and training techniques can reduce the problem.
Claude Code pulled from Pro, now costs individuals 5x more
Anthropic has removed Claude Code from its $20/month Pro subscription, calling it a 'test' despite updating its pricing page. Journalist Ed Zitron spotted the change on April 21st. The coding assistant remains available on Team plans, which require a minimum of 5 seats at $20 each, making the cheapest path $100/month. Users have also reported severe throttling issues with Claude Opus recently.
Trainly's free audit found $2,400/mo in wasted GPT-4 calls
Trainly offers a free 72-hour AI agent observability audit that analyzes production traces to identify cost concentrations, blind spots, and drift alerts. The audit uses heuristics to detect silent tool failures, retry loops, latency and cost outliers, error patterns, and behavioral anomalies without making LLM calls on their end.
Kuri's 464KB browser agent beats Vercel's on token cost
Kuri is a browser automation and web crawling tool written in Zig, designed specifically for AI agents with a focus on token efficiency. It features a 464 KB binary with ~3ms cold start, and claims 16% lower workflow token cost compared to Vercel's agent-browser according to project benchmarks. It offers CDP automation, accessibility snapshots, HAR recording, a standalone fetcher mode (kuri-fetch) that doesn't require Chrome, and an interactive terminal browser.
OpenAI's Workspace Agents Come for Zapier's Lunch
OpenAI shipped workspace agents for ChatGPT, putting it in direct competition with Zapier and UiPath. These shared, Codex-powered agents handle complex tasks and long-running workflows within organizational permissions. They work autonomously across tools and teams, run on schedules or in Slack, and are available for Business, Enterprise, Edu, and Teachers plans. The feature includes enterprise governance, approval controls, and reusable team workflows.
Show HN submissions tripled and now mostly have the same vibe-coded look
Analysis of 500 Show HN pages reveals 67% show signs of AI-generated design patterns, with 21% exhibiting 'heavy slop' (5+ patterns) and 46% showing 'mild' patterns (2-4). The surge in generic, AI-assisted designs is attributed to tools like Claude Code, which has led to a threefold increase in submissions and prompted HN moderators to restrict Show HN for new accounts. Common AI design patterns include Inter fonts, 'VibeCode Purple', shadcn/ui components, glassmorphism, centered hero sections, colored border cards, and gradient backgrounds.
MythosWatch Reveals Who Gets Anthropic's Restricted AI
MythosWatch monitors access to Anthropic's restricted Claude Mythos Preview model, aggregating 39 public signals across named access, reported use, and policy negotiations. Current data shows 12 named partners and 40+ undisclosed organizations, with early access concentrated in infrastructure, security, finance, and government. A Bloomberg report revealed unauthorized access through a third-party vendor. Distribution patterns follow Five Eyes alignment with notable EU gaps.
Qwen3.6-27B beats Claude Opus 4.5? Benchmark methods questioned
Qwen3.6-27B is a new open-weight 27B parameter large language model released on Hugging Face with multi-modal capabilities (image-text-to-text), tool calling/function use, and reasoning support. Benchmarks show strong performance on MathArena AIME 2026 (94.1) and GPQA (87.8). HN commenters note it reportedly beats Claude Opus 4.5 in internal benchmarks, though questions are raised about non-standard Terminal-Bench 2.0 methodology.
Martin Fowler: Coding Agents Need More Laziness
Martin Fowler argues that good developers are lazy in useful ways, building abstractions to avoid repetitive work. LLMs lack this virtue and produce bloated code instead. He explores applying TDD to agent prompting and argues that teaching AI when NOT to act is essential for safe autonomous systems.
AI Was Ruining My Philosophy Class. So We Wrote One Essay Together
When AI made traditional philosophy essays unreliable, a University of Chicago professor tried something unusual: writing one with his entire class. The collaborative experiment worked. Students said they worked harder, learned more, and were finally doing real philosophy instead of pretending for a grade.
SpaceX Bets $60B That Cursor's AI Edge Outweighs Its China Problem
SpaceX has agreed to buy AI code editor Cursor for $60 billion, per Bloomberg. The deal includes a $10B breakup fee and raises immediate questions about Chinese AI dependencies in a company handling classified US payloads.
Flipbook.page: AI Builds Websites Live On Request
Developer Daryl's project streams website content live from Google's Gemini API, generating web code in real-time. The demo faces rate-limiting issues and performance limitations with current GPU speeds.
Dead Startups Are Selling Your Slack Messages to Train AI
Defunct startups are selling internal data including emails, Slack messages, and source code for $10,000-$100,000 to train AI models. SimpleClosure has processed nearly 100 such deals in the past year. The data feeds 'reinforcement learning gyms' where AI agents practice workplace tasks. Privacy advocates say employees never consented to this.
This Developer Writes Half His Code by Hand. On Purpose.
marcgg writes at least half his production code by hand. Not because he's anti-AI, but because he's watched what happened to pilots who forgot how to fly. His workflow splits the difference: vibe code the toys, handwrite the stuff that matters.
Qwen3.6-27B Rivals Claude on Code, No Cloud Required
Qwen3.6-27B is a new open-weights 27B parameter dense model focused on coding, with early users reporting strong results in C, C++, and Verilog. The model improves on Qwen3.5-27B and positions itself as a local alternative to cloud-based options like Claude, particularly for developers working with proprietary code or hitting usage caps.
Zed's Parallel Agents: Multiple AIs, One Editor Window
Zed editor ships Parallel Agents, letting developers run multiple AI agents simultaneously in one window. With a new Threads Sidebar and redesigned layout, it offers multi-agent orchestration that Cursor and VS Code don't currently match.
PI Dashboard gives you live control over running agent sessions
PI Dashboard is an open-source web tool that lets you monitor coding agent sessions in real time and jump in while they're still running. You can send prompts, kill runaway agents with escalating force, and manage multiple sessions across projects from any browser.
Medievalizer turns docs into illuminated manuscripts
Medievalizer is a Chrome extension that transforms documentation pages into illuminated medieval manuscripts with blackletter headings and Shakespearean prose. Powered by Claude Sonnet, it preserves code blocks and technical accuracy while converting prose into archaic language with features like streaming output, drop caps, and one-click restore functionality.
Google's Eighth-Gen TPUs Split Into Training and Inference Chips
Google announced its eighth-generation TPU chips, now split into specialized processors for training and inference workloads. The TPU 8i inference chip features 384MB of SRAM (triple the previous generation) and is designed to deliver the throughput and low latency needed to run millions of AI agents. Companies including Citadel Securities, U.S. Energy Department national laboratories, and Anthropic are already using Google's TPUs.
Aide replaces Google Assistant with any AI model you choose
Aide is a BYOK AI assistant for Android that replaces Google Assistant. Connect to Claude, ChatGPT, Gemini, local Ollama, or any OpenAI-compatible endpoint. Features include voice input/output, device actions (SMS, calls), Home Assistant integration, and web search. API keys encrypted on-device. Free with a $9.99 one-time Pro upgrade.
The zero-days are numbered
Mozilla used Anthropic's Claude Mythos Preview to find 271 security vulnerabilities in Firefox 150. Opus 4.6 previously caught 22 bugs in Firefox 148. The 12x increase in detection raises questions about whether AI is shifting the advantage from attackers to defenders.
Passive Contact Lens Treats Glaucoma Without Electronics
Researchers at the Terasaki Institute for Biomedical Innovation have developed a smart contact lens with zero electronics that uses microfluidics to monitor glaucoma and automatically deliver drugs when eye pressure climbs. A smartphone app with a convolutional neural network reads pressure levels with 94% accuracy. Testing on rabbits showed effectiveness comparable to eye drops with no biocompatibility issues over 14 days of use.
Anthropic's Mythos Model Leaked to Unauthorized Users on Launch Day
A private online forum group gained unauthorized access to Anthropic's Mythos model the same day it was announced for limited testing. The group has used the model regularly since, corroborated by screenshots and a live demo.
Hydra: Swap AI coding CLIs mid-session when rate limits bite
Hydra is a unified wrapper for AI coding CLIs that switches between providers like Claude Code, Codex, OpenCode, and Pi when hitting rate limits. It automatically manages context transfer through clipboard, letting developers maintain workflow without manual context copying or re-explanation.
Mozilla Squashed 271 Firefox Bugs Using Anthropic's Mythos
Mozilla used Anthropic's Mythos Preview AI model to identify and fix 271 vulnerabilities in Firefox 150, gained through direct collaboration with Anthropic. The find shows AI can now catch bugs that previously required expensive human analysis. But the approach raises questions about access: most open source projects lack the resources and connections that made this possible.
Taskd lets Claude manage its own task queue
A task management system built by Levi Durfee that integrates with Claude AI via an MCP server. Written in Go and TypeScript, deployed on Google Cloud Run with Cloud SQL and envelope encryption via GCP KMS. Claude can autonomously pick up tasks, create plans, leave comments, change status, and even add suggestions to the task list.
Continuous-Refactoring Spends Spare API Tokens on Code Cleanup
An open-source tool runs AI agents in a loop: propose a refactor, run tests, keep only changes that pass. Started as a simple janitorial loop, then added taste preference files and a staged migration workflow for bigger jobs.
SpaceX buys Cursor for $60B. Users' code may be the real prize.
SpaceX is acquiring AI code editor Cursor for $60B. The real prize is the massive repository of developer code that could feed xAI's training pipeline.
Anthropic's "Stolen" Mythos Model: Real Breach or Hype?
Financial Times reports Anthropic is investigating unauthorized access to a model called Mythos. The article is behind a paywall with limited details. Hacker News commenters question whether the story is genuine security news or clever marketing.
Mozilla Let Anthropic's AI Loose on Firefox. It Found 271 Bugs.
Firefox 150 shipped this week with fixes for 271 security bugs found entirely by Anthropic's Mythos AI. Automated tools can now catch what previously required human analysis. Mozilla says that's good news for Firefox users, but signals a rough transition for everyone else.
Anthropic hikes Claude Code to $100/month as quality drops
Anthropic removed Claude Code from the $20/month Pro plan and now requires a $100/seat/month Team Premium seat. The change coincides with documented quality regression tied to a February update, with AMD engineer Stella Laurenzo's analysis of over 6,800 sessions showing the assistant began ignoring instructions and hallucinating fixes. Users on Hacker News expressed frustration, with some considering competitors like GLM and Kimi or exploring local models.
Meta's Plan to Turn Employees Into AI Training Data
According to reports, Meta plans to capture employee mouse movements, keystrokes, and screen activity to train AI agents. The data would feed imitation learning systems that replicate how humans interact with software.
Zindex Wants to Be the Database for Agent Diagrams
Zindex is a diagram infrastructure platform designed for agents and agentic systems, featuring the Diagram Scene Protocol (DSP) that enables agents to create, edit, validate, and render diagrams as durable state. Key features include semantic descriptions (not geometric), built-in Sugiyama-style hierarchical layout pipeline, incremental editing with stable IDs, multiple render targets (SVG, PNG with 4 themes), deterministic execution, 40+ validation rules, and PostgreSQL storage. It serves as the middle layer between agent reasoning and visual output.
Meta logs workers' keystrokes and clicks to train AI agents
Meta is installing tracking software on US employees' computers to capture mouse movements, clicks, and keystrokes for AI model training. The data will help build agents that can perform computer tasks autonomously, learning how humans interact with dropdown menus, keyboard shortcuts, and other interface elements. Spokesperson Andy Stone says the data won't be used for performance reviews.
Codemix ships a graph DB that type-checks queries and syncs via CRDTs
@codemix/graph is an open-source TypeScript graph database built on CRDT technology, featuring type-safe schema definition with Zod/Valibot/ArkType, Gremlin-style traversal API, Cypher-like query language support, and Yjs-based offline-first collaboration. It enables LLMs to execute structured queries against graph data.
The blurry JPEG gets a name: expansion artifacts
An opinion piece introducing 'expansion artifacts,' the term for hallucinations, style issues, and strange outputs that appear when LLMs generate content. Unlike compression artifacts, these occur during decompression, when models extrapolate from compressed training data. The article examines examples in text, code, images, and video, and warns of risks when AI-generated content feeds into new AI generations, creating feedback loops that flatten and worsen information quality.
gpt-image-2 drops diffusion for transformers, tops image arena
ChatGPT launches Images 2.0, an upgraded AI image generation feature with advanced text-heavy composition capabilities, expanded editing tools, high-resolution outputs up to 4K, and flexible format support for infographics, posters, comics, and social content.