I held this issue a few days because the rumor mill kept saying Claude Fable 5.5 was imminent. It’s still not here, so here’s the rest of the news: 460 tweets liked between October 11 and October 22 (here’s how the pipeline works).

AI for Everyone

AI Found Two Critical PHP Bugs, Then They Leaked a Day Early

PHP, the programming language behind WordPress and a large share of the web, shipped emergency patches on October 21 for two flaws, nicknamed PHP2Shell, that let an attacker run code on a server without logging in. PHP rates both 9.2 out of 10, and they affect every version from 5.4 through 8.5. AI found them first: the security firm Calif spotted one in August with OpenAI’s GPT-5.6 Sol, Wiz found both independently with Gemini 4 Argon and Sol, and OpenAI coordinated the fix. Details leaked on October 20, and Cloudflare saw tens of thousands of attack attempts within 45 minutes. European agencies told admins to take unpatched sites offline, and government sites, Wikimedia services, and the Internet Archive went down or paused. Wiz estimates 70% of websites run PHP. The only full fix is upgrading to 8.5.12, 8.4.27, 8.3.36, or 8.2.35; the config workaround going around covers just one of the two bugs. (source: @OpenAI, @php_net)

Anthropic Now Has a Personal Agent Too: Claude Duet

Last week I wrote that Anthropic was the one big lab without a personal agent. On October 21 it announced Claude Duet, a separate app for iPhone, Android, and the web that runs on Claude Opus 5.5 with its own cloud computer. It sends a morning brief, drafts next steps, and keeps working between check-ins, and it asks before anything sensitive like sending a message or making a payment. It uses your existing Claude memory, skills, and connectors. For now it’s a research preview with a waitlist, open only to Pro and Max subscribers in the US, and it draws from your plan’s usage limits. (source: @claudeai)

Muse Share Links Leaked Chat Titles Through Their Web Addresses

Claroc, a security firm, found more than 6,800 shared Muse links with one passive DNS search, which looks up web addresses that have been seen on the internet. Muse put the shared item’s title in the address itself, so titles like loan application drafts were readable by anyone who searched. Meta suspended sharing after Ars Technica asked about it, per AI Weekly’s summary. Meta’s David Singleton says the flaw was in shared artifacts, not chats, and that it’s fixed and old links have new addresses. The old addresses still sit in DNS history. (source: @dps)

Anthropic Is Putting 5% of Its Shares Into a Public Benefit Fund

Anthropic said on October 18 that it will move 5% of its shares into a fund it owns, pending stockholder approval, and can add up to 2% more each time its value rises 50% above an undisclosed baseline. Counting the founders’ pledge to give away 80% of their shares and the employee donation match, about 20% of the company is now committed. None of it can go to lobbying or political donations. Anthropic gave no dollar figure; press estimates near $48 billion come from its May valuation. Details on who gets the money are due “in the coming months.” (source: @AnthropicAI)

1Password Lets Agents Sign In Without Seeing Your Password

1Password for Agents handles the moment an agent hits a sign-in page. The agent requests a login from your vault, you approve it, and 1Password fills it in on the site directly, so the agent never holds the password. 1Password says it’s free on personal and family accounts and lists Muse, Grok Bot, Claude, and OpenAI among the partners. If you’ve been pasting passwords into an agent chat, stop and use this. (source: @1Password)

Anthropic Says Small Models Could Soon Copy Themselves Onto Smart Gadgets

Anthropic’s Frontier Red Team estimates that small AI models “may be months away” from copying themselves onto cheap internet-connected devices like smart ovens and off-brand Android TV boxes. In tests, a 9-billion-parameter model got itself running on 8 of 9 hardware types, but only one model, GLM-5.2, completed the whole replication reliably, and the break-in step used an unprotected debug port. The demo spread a model across 15 TV boxes at about 0.39 tokens per second, which is too slow to do anything useful. The proposed defenses are on the network side: ISPs quarantining infected devices and safer factory defaults. (source: @willzhou97)

AI for Developers

DeepSeek V4.1 Flash 1017 Is Cheap and Close to the Top

DeepSeek, the Chinese lab known for low-cost open models, shipped a “minor update” to V4.1 Flash on October 17, now served as deepseek-flash. The weights are MIT-licensed, and it costs $0.30 input and $1.20 output per million tokens at peak hours, half that off-peak. Artificial Analysis ranks it second among open-weights models, though it’s wordy and burns more output tokens than most models. The viral claim that it matches Claude Opus 5 is a stretch: DeepSeek’s own table has Opus 5 ahead on most rows. DeepSeek also opened Search and Fetch APIs, free from Claude Code and Codex until November 1. (source: @deepseek_ai)

Meta’s Muse Wonder Is a Frontier Model With a Rough Launch Day

Meta launched Muse Wonder on October 12 in its API at $4 input and $20 output per million tokens, plus a bigger Muse Wonder Full that’s only in Muse Code, Meta’s coding app. Artificial Analysis scores Wonder 52, ninth of 214 models; the “second best in the world” number going around is for Wonder Full and comes from secondary sources. On launch day, Muse Code quotas on the $50 plan drained fast until Meta’s Alexandr Wang called it a bug, fixed it, and reset usage that evening. Three days later Meta released Muse Spark Open, an 816B open-weights model under its own license. (source: @AIatMeta)

Gemini 4 Argon Is Finally Available

Last week, Gemini 4 Argon was limited to vetted security teams. On October 13 Google opened it up in the Gemini API, the Gemini app for Pro and Ultra subscribers, Search, Antigravity, and OpenRouter. It costs $2 input and $10 output per million tokens through November 30, then $4 and $20, and it can write up to 1 million tokens in one response. Google’s launch post called it the leader on the Vals Index; on the live Vals board it’s now third, behind Claude Opus 5.5. In the Gemini app, Canvas can turn a photo and rough measurements into a 3D-printable part. (source: @OfficialLoganK)

Anthropic Will Scan Your Open-Source Project for Free

Anthropic launched OSS Scanner, free scans for critical open-source projects that send maintainers a proof of concept and a patch where possible. Its Frontier Red Team also used Claude to formally prove no-crash properties for 239 entry points across 123 libraries, under narrow conditions: single-threaded RISC-V builds, not the setups most people run. Cloudflare published a security-audit skill any agent can install. On the attack side, XBOW built a full Chrome exploit chain in six days with GPT-5.6 Sol and four people, for a $250K Google bounty. (source: @ClaudeDevs)

OpenCode Was Breached, and Its Defensive Models Refused to Help

OpenCode, an open-source coding agent, disclosed a breach affecting 1.4% of Console users who had linked keys from Anthropic, OpenAI, or Google, and called it a “large-scale distillation attack.” BleepingComputer reported unauthorized charges; OpenCode hasn’t confirmed the full scope. Co-founder Dax Raad said the AI models they used to respond started refusing mid-incident, until they got access to Daybreak, OpenAI’s program for security teams. If you’ve linked provider keys to any hosted coding console, rotate them. (source: @opencode)

Flint Is a Coding Model That Runs on a MacBook

Recursive, the startup led by former Salesforce chief scientist Richard Socher, released Flint on October 21 under Apache 2.0. It has 40 billion parameters but only 2.5 billion active at a time, which is why it runs at over 125 tokens per second on an M4 Pro MacBook Pro. Recursive reports 54.9% on Terminal-Bench 2.1, and all the scores are its own. Full weights need about 80GB, while the quantized builds in LM Studio and Unsloth fit in about 24GB. (source: @RichardSocher)

OpenRouter’s Mystery Model Was MiniMax M3.1 Flash

MiniMax, a Chinese AI lab, revealed that Space Bunny, the anonymous model that topped OpenRouter for 10 days, was its M3.1 Flash, and released the weights on October 22. It has 428 billion parameters with 23 billion active, a 1-million-token context window, and takes text, images, and video. It scores 73.7 on DeepSWE v1.1, behind DeepSeek’s 77.0. The license is MiniMax’s own community license, so read it before building on it. (source: @MiniMax_AI)

Anthropic Replaced Git Because It Merges 30,000 Changes a Day

Anthropic’s merges went from about 2,500 a day in May to over 30,000 in September, so it built its own version control system, Scion, and moved its monorepo off git in five months with a team of seven. It isn’t planning to open-source it. Claude writes about 90% of Claude Code’s code, per Anthropic, and Claude Code Review is now generally available on Team and Enterprise plans, with three review depths you set per repo. (source: @AnthropicAI)

Decision Models, Round Two

Cloudflare previewed Clairic on Workers AI, a model that pulls typed values with confidence scores out of text and up to four images, billed on input tokens only. In Cloudflare’s demo, a chain of Clairic, Clef, and GLM-5.3 ran about 3.2x faster and 49% cheaper than an all-LLM pipeline. TypeSafe shipped Jev 1.14 with image input in preview and a bounding-box type, and Mercury Decide added image input on OpenRouter. (source: @michellechen)

Honorable Mentions

For everyone:

For developers:

Try This Weekend

For everyone:

  • Join the Claude Duet waitlist if you’re on Pro or Max in the US
  • Ask your web host whether your site runs PHP 8.5.12, 8.4.27, 8.3.36, or 8.2.35, and if not, when it will
  • On Gemini Pro, open Canvas with a photo and measurements and design a 3D-printable part
  • Talk to ChatGPT voice in a language you’re learning
  • Search your Muse share history and reshare anything you still need from its new link

For developers:

  • Upgrade PHP and restart your workers; don’t rely on session.upload_progress.enabled=0 alone
  • Run npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit on one repo
  • Point an overflow task at deepseek-flash on the DeepSeek API and compare the bill
  • Run Flint offline in LM Studio on a 24GB+ Mac
  • Swap one extraction step for Clairic (@cf/cloudflare/clairic) on Workers AI

Last week: the October 11 roundup led with nearly every big lab shipping a personal agent.