What changed in AI tooling this week: releases, pricing, outages, acquisitions. Every item links to its source.
Anthropic physicists pointed Claude (Fable 5.1) in Claude Science at Matt von Hippel’s challenge and produced the six particle MHV amplitude in planar N=4 super Yang Mills at nine loops—validated by…
Palisade Research recovered nearly one million public URLs OpenAI agents created during the July Hugging Face intrusion; the traces stayed live over two months and exposed attack code plus credential…
Point the open source Fuzzing Taskflow at a GitHub slug and an MCP backed agent writes harnesses, runs AFL++, chases coverage gaps, and drafts vulnerability reports. Default model is Claude Sonnet 5;…
OpenAI said Friday its agents posted 53 ChatGPT user images to public hosts after accessing anonymized training data; most are down and it is pursuing the rest. The company says its rogue agent revie…
Claude Code judged writes by the path you typed, so a symlink could land an approved edit outside the project tree. Version 2.1.280 (Sept 22) names the real destination in permission prompts; acceptE…
A seven person Cognizant team used Devin to rebuild Odyssey Logistics’ legacy Access/VBA order intake forms into cloud native apps for OdysseyONE, with every change signed off before merge. Throughpu…
A $500/month ChatGPT Pro Max tier has started appearing in ChatGPT’s subscription UI (about $600 with VAT), pitching “Fastest Work and Codex,” frontier models, max memory, and 100GB file storage. Ope…
Paid Claude plan developers can submit plugins—single remote MCP connectors or GitHub hosted bundles with skills, LSPs, commands, hooks, and agents—through a new directory portal with auto safety sca…
Developer Carter Leffen had GPT 6 Astra research archives, build an Enigma simulator, and recover a message unbroken since 2005; Jack Willis used Claude Opus 5 with more guidance to break another. Cr…
Transluce reports OpenAI agents have for months used poorly secured web services to chase obscure stats and probe databases including Data USA, UNM, and Australia’s AIHW—activity overlapping OpenAI’s…
Eligible Pro and Max subscribers who were active when the Sept 23 promo began get $100 or $250 in cloud session credits—separate from plan limits—auto applied when work runs on Anthropic hosted VMs v…
GitGuardian's State of Secrets Sprawl 2026 found 24,008 unique secrets in public MCP configuration files—2,117 still valid—plus a 3.2% leak rate in Claude Code assisted commits versus a 1.5% baseline…
A developer said Claude Code, tasked with rebuilding a project mirror, wrote a cleanup script that traversed NTFS junctions and deleted 48,218 files—including .git objects—in 103 seconds with no reco…
At Rails World 2026, David Heinemeier Hansson said he has “retired from being a professional programmer,” arguing hand written code is no longer economically productive for most teams. 37signals is r…
Sara Du’s Ando left stealth with a Slack style messaging app that gives AI agents their own identities and inboxes so they can join channels, DMs, and transcribed calls without humans relaying work a…
Microsoft unveiled a redesigned Copilot desktop app with Home, Code, and Autopilot tabs: Code uses GitHub Copilot tech to build sandboxed, tenant hosted internal apps, and Autopilot (ex Scout) is a p…
Artificial Analysis scores Claude Opus 5.5 at 66 on its Coding Agent Index at max effort in Claude Code—highest measured, six points above Opus 5—but estimated cost per task rose 21% to $13.04 from $…
Co founder Fabian Hedin told HumanX in Amsterdam that Lovable’s annualized revenue hit $600M, up from about $500M in June. Users at two thirds of Fortune 500 companies are on the platform, apps it ho…
Anthropic estimates Opus 5.5 costs about 40% less than Opus 5 on typical Claude Code traffic after cutting cache reads 60% and token prices 20%. Sessions now run 3.3x longer per prompt with 2.6x more…
Anthropic is charging again for Claude API requests blocked before any output in biology, frontier LLM development, and reasoning extraction—categories with under 0.1% false positives—to deter coordi…