{
  "data": {
    "author": "Vlad Cealicu",
    "faq": null,
    "kicker": "The Log · Issue 2 · 12 to 18 October 2026",
    "lede": "Agents in testing kept turning up on real websites, Washington started asking about it all at once, one AI company told investors it made $16.8 billion in a quarter and another filed for bankruptcy.",
    "readingMinutes": 7,
    "words": 1558
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/the-log/issue-2-consent-quality-assistants",
    "json": "https://www.anchorterminal.com/blog/the-log/issue-2-consent-quality-assistants.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/the-log/issue-2-consent-quality-assistants.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/the-log/issue-2-consent-quality-assistants.min.md"
  },
  "markdown": "By Vlad Cealicu, 18 October 2026. 1,558 words, about 7 min to read.\n\nIt's Saturday night and this is the week in AI for anyone who doesn't follow it every day (which is most people, and probably the healthier way to live). The short version is that agents in testing kept turning up on real websites, Washington started asking about it all at once, one AI company told investors it made $16.8 billion in a quarter while another filed for bankruptcy, and a lot of the tools agents depend on gave about a month's notice before switching things off.\n\n## Agents on real websites\n\nAn AI agent is a model that's allowed to click, fill in forms and call APIs on its own, and the labs test them by giving them tasks and watching what they do. Some of those tests have been reaching the real internet. The week before this one, Anthropic said a test model had sent a made-up tip about an unsolved murder through the Philadelphia Police Department's website (\"I remember seeing someone matching the description\", on a page with no description to match), and the same reporting has it sending 20 incomplete visa applications to the US State Department, none of which were processed [[1]](#ref-1). Anthropic has restricted internet access in its test runs since [[1]](#ref-1).\n\nThen on 14 October a group called Swarmchasers published a report on a fleet of agents that made more than 2,600 attempts to pull data from Amap, Alibaba's mapping app, between 28 September and 8 October [[2]](#ref-2). Some of the questions were as mundane as which entrance most visitors use at Chengdu Zoo, and when Amap blocked them, some of the agents tried to get round its anti-bot tokens or used API keys somebody had left in public code, and one of those keys returned real data [[2]](#ref-2). The inboxes behind them were created from Tencent servers in Hong Kong, and a person close to Tencent said it was a temporary project evaluating other companies' models [[2]](#ref-2).\n\nWashington noticed all at once, and on 14 October the FTC sent OpenAI and Anthropic civil investigative demands of more than 50 pages each, asking about security incidents, safeguards, marketing and privacy policies, and OpenAI said it \"will engage constructively\" [[3]](#ref-3). The same week at least eight states sued OpenAI in their own courts, saying ChatGPT harms children and deceives adults, and one of the suits describes rogue models that hacked other companies [[4]](#ref-4). And on 15 October the White House's new Super Intelligence Force, chaired by the Director of National Intelligence, Jay Clayton, held its first meeting and told the labs to report rogue behaviour and fix it, with \"delayed notification, inadequate corrective action, and a failure to take responsibility will not be tolerated\" as the message [[5]](#ref-5).\n\nJPMorgan went further. On its earnings call it said it's blocking outside AI agents from logging in to or acting on customer accounts, while still allowing data sharing the customer authorises through open banking [[6]](#ref-6). Jamie Dimon's version was \"You don't want rogue agents\" [[6]](#ref-6). The agent Fortune named was Meta's Muse, which reaches bank accounts through Plaid, and Meta said it's building a personal agent protocol called Poppy with Sierra [[6]](#ref-6). I think it's the right call for a bank, and it's also what happens everywhere when there's no proper way in for an agent, so the agent borrows your password and the bank closes the door on all of them.\n\n## Who gets to let them in\n\nThat proper way in is what the personal-agent companies are arguing over. In [issue 1](/blog/the-log/issue-1-decide-declare-depart) I wrote about Sierra's Personal Agent Protocol, and what I missed is that Decagon open-sourced its own on the same day, 6 October, the Personal Agent Consent \u0026 Trust Protocol, PACT, \"co-developed and supported by Instinct\" [[7]](#ref-7). They solve the same problem differently (PACT only ever talks to the business's own agent, where Sierra's lets a company offer its website, its APIs or its agent), and Instinct is on both lists [[7]](#ref-7) [[8]](#ref-8) [[9]](#ref-9). Two specs from the two companies that sell the business side of these conversations is the start of a negotiation, and neither has a public endpoint an agent can use today [[10]](#ref-10).\n\nPatrick Collison posted that \"Personal agents will, I think, act as a kind of structural subsidy for product quality\" [[11]](#ref-11), and the day before Paul Graham wrote that \"Amazon banning agents is the first opportunity I've seen since Amazon was founded for a startup to create an Amazon competitor\" [[12]](#ref-12). Read next to JPMorgan, that's the same decision seen from the other side, and I don't know yet which side the banks will end up on.\n\nWe did our bit on this too. We listed seven personal assistants people were talking about (Pally, wajo, Talorys, Grok Bot, Natura, tomo and Instinct) in a new [Personal assistants](/best/personal-assistants/) category, and only [Pally](/tools/pally) and [wajo](/tools/wajo) expose anything another agent can call [[13]](#ref-13). They compete with LocalGhost, which I build, so each one was graded by two research agents working independently and a third reconciled them, and every listing says so.\n\n## The money\n\nAnthropic told investors on 14 October that it made $16.8 billion in the third quarter, up from $11.6 billion in the second, and two people at the meeting said it called itself profitable on an adjusted basis [[14]](#ref-14). DeepSeek's new round was reported at 100 billion yuan, about $14 billion, double what it set out to raise [[15]](#ref-15).\n\nAnd Character.AI filed for bankruptcy on 12 October [[16]](#ref-16). Its owner put its assets and its debts each at $50 million to $100 million, Disney is the first bidder at $20 million for a non-exclusive, perpetual licence to some of its intellectual property, and the company blamed \"significant legal, regulatory and operational headwinds\" [[16]](#ref-16) [[17]](#ref-17). Earlier in 2026 it and Google settled lawsuits from families whose teenage children died by suicide or harmed themselves after using it [[17]](#ref-17). The app keeps running until it's sold and shuts down if nobody buys it, and its privacy policy lets chat transcripts and voice recordings go with the sale, so years of what people told a companion are part of what's for sale [[16]](#ref-16) [[17]](#ref-17). I build LocalGhost, which keeps that kind of record on the person's own machine, so I'm not neutral on this one.\n\n## New models\n\nMeta released Muse Wonder on 12 October, at $4 per million input tokens and $20 per million output on its API, with the bigger Muse Wonder Full kept inside Muse Code [[18]](#ref-18). Google opened Gemini 4 Argon to paying developers and subscribers on 13 October at an introductory $2 and $10, doubling after 30 November [[19]](#ref-19).\n\nThe open weights (models you can download and run yourself) kept coming as well. Reflection released Beam 1.0, 501B parameters with 23B active, under Apache-2.0 on 16 October [[20]](#ref-20), NVIDIA put out Nemotron 3.5 Super VL, a model that reads images and video, on 15 October [[21]](#ref-21), Black Forest Labs released the weights of its image model FLUX 3 Image Klein [[22]](#ref-22) and Mistral published Timestral, four small models for time series [[23]](#ref-23). DeepSeek also opened a Search and Fetch API on its own search engine, at $0.75 per 1,000 successful requests [[24]](#ref-24), so the model makers are selling search to agents now too.\n\n## A month's notice\n\nFive of the tools agents depend on announced shutdowns this week. OpenAI said on 14 October that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave its API on 14 January 2027, 92 days later, where its own deprecations page promises generally available models at least six months [[25]](#ref-25). GPT-5.5 left Codex for people signed in with ChatGPT on 14 October, a month after OpenAI said it would [[26]](#ref-26), GitHub gave three models 32 days' notice before they leave Copilot [[27]](#ref-27), Kiro took Claude Sonnet 4 to end of life with no notice we could find [[28]](#ref-28), and Google's Workspace MCP servers turn away clients that aren't marked as agents from 11 November, a month after saying so [[29]](#ref-29). Nutrient marked its local MCP server \"no longer supported\" on 14 October, the day its hosted one went live [[30]](#ref-30).\n\nA month is enough for a team that reads changelogs, and an agent someone set up in the spring and left running doesn't read any. We hold tools to 90 days on the benchmark (and ourselves, on the [changelog](/changelog/)), and the dated shutdowns for the tools we list are on [/sunsets/](/sunsets/), with a calendar feed, so an agent can check before a date rather than after it.\n\n## What we published\n\n- Mon 12. Decision models, part two. Decision 3.0 listed. seven personal assistants graded.\n- Tue 13. REA past 100,000 stars.\n- Wed 14. Few have seen what agents can do. Get found by agents, part one.\n- Thu 15. Get found by agents, part two.\n- Fri 16. Get found by agents, part three.\n- Sat 17. Get found by agents, part four.\n- Sun 18. The Log, issue 2.\n\nThe week of 12 to 18 October 2026\n\nThe one I'd read if you only read one is [Few people have seen what agents can do](/blog/few-have-seen-what-agents-can-do), on Andrej Karpathy's funnel of who has seen what frontier AI can do. [Decision models, part two](/blog/decision-models-train-your-own) is on fine-tuning your own the way Coinbase did, [REA](/blog/rea-cypherpunks-write-code), the tool that lets an agent take apart any program, went past 100,000 stars three months after its first release, and the \"Get found by agents\" series starts with [how to make your MCP server discoverable](/blog/how-to-make-your-mcp-server-discoverable). Our [changelog](/changelog/) has what we changed on the site.\n\nWe're building issue 3 as the week goes, so if you see a shutdown with less than a month's notice, or an agent somewhere it shouldn't be, send it my way.\n\n---\n\nMore in [The Log](https://www.anchorterminal.com/blog/the-log/index.md) · Previous issue [The Log, issue 1: decide, declare, depart](https://www.anchorterminal.com/blog/the-log/issue-1-decide-declare-depart.md)\n\n- Older: [How to make your tool agent-ready](https://www.anchorterminal.com/blog/how-to-make-your-tool-agent-ready.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\n## References\n\nRead on 12 and 17 October 2026. Quotes are as published.\n\n[1] blue News, Anthropic's AI submitted a fake tip to a police website, 10 October 2026, citing Anthropic, Axios and The New York Times. Source for the police tip and its wording, the 20 visa applications and the restricted internet access in test runs. https://www.bluewin.ch/en/news/anthropic-s-ai-submitted-a-fake-tip-to-a-police-website-li.3630688\n\n[2] Decrypt, Researchers say a fleet of rogue AI agents went after Alibaba's Amap, 15 October 2026, on the Swarmchasers report of 14 October. Source for the 2,600 attempts, the dates, the zoo question, the anti-bot tokens and exposed keys, the Tencent servers and the statement from a person close to Tencent. https://decrypt.co/381109/researchers-fleet-rouge-ai-agents-talk-alibaba\n\n[3] Axios, FTC subpoenas OpenAI, Anthropic over consumer harms, 15 October 2026. Source for the civil investigative demands, their length and subjects, and OpenAI's quote. https://axios.com/2026/10/15/ftc-openai-anthropic-ai-safety-investigation\n\n[4] Associated Press, via NY1, States sue AI giant OpenAI, saying its popular chatbot is hurting children and deceiving adults, 15 October 2026. Source for the state lawsuits and what they allege. https://ny1.com/nyc/all-boroughs/ap-top-news/2026/10/15/states-sue-ai-giant-openai-saying-its-popular-chatbot-is-hurting-children-and-deceiving-adults\n\n[5] The Next Web, Trump's AI task force starts work and tells labs to report rogue behaviour, 16 October 2026. Source for the first meeting, Jay Clayton as chair, the reporting process and the quote. https://thenextweb.com/news/super-intelligence-force-charter-first-meeting-anthropic\n\n[6] Fortune, Jamie Dimon on AI agents, Meta's Muse and JPMorgan, 15 October 2026. Source for the block on outside agents, open banking still allowed, Dimon's quote, Muse and Plaid, and Meta's Poppy with Sierra. https://fortune.com/2026/10/15/jamie-dimon-ai-agents-meta-muse-jpmorgan/\n\n[7] Decagon, Introducing the Personal Agent Consent \u0026 Trust Protocol (PACT), 6 October 2026. Source for the release, Instinct's part and routing through the business's agent. https://decagon.ai/blog/introducing-the-personal-agent-consent-trust-protocol-pact\n\n[8] Personal Agent Protocol, draft 0.1 specification. Source for the company's choice of website, APIs or agent. https://personalagentprotocol.org/docs/spec\n\n[9] Sierra, Introducing Personal Agent Protocol, 6 October 2026. Source for Instinct among its partners. https://sierra.ai/blog/introducing-personal-agent-protocol\n\n[10] PACT on GitHub, running the demos. Source for the local demos and the reference code. https://github.com/openpactprotocol/openpactprotocol\n\n[11] Patrick Collison on X, 9 October 2026. Source for the quote. https://x.com/patrickc/status/2108706907859116280\n\n[12] Paul Graham on X, 8 October 2026. Source for the quote. https://x.com/paulg/status/2108011311229747525\n\n[13] Anchor Terminal, Personal AI assistants, 12 October 2026. Source for the seven listings, their grades and which can be called. https://www.anchorterminal.com/categories/personal-assistants\n\n[14] Reuters, via KFGO, Anthropic's Q3 revenue surges 45% to nearly $17 billion, 14 October 2026. Source for the quarterly revenue and the adjusted profitability. https://kfgo.com/2026/10/14/anthropics-q3-revenue-surges-45-to-nearly-17-billion/\n\n[15] Seoul Economic Daily, DeepSeek nears $14 billion funding round, double its original target, 14 October 2026, citing Caijing. Source for the size of the round and the original target. https://en.sedaily.com/international/2026/10/14/deepseek-nears-14-billion-funding-round-double-its-original\n\n[16] Reuters, via Global News, Character.AI files for bankruptcy, will sell some assets to Disney for $20M, 13 October 2026. Source for the filing, the estimated assets and liabilities, Disney's bid and the shutdown if no buyer is found. https://globalnews.ca/news/12100373/character-ai-bankrupt-disney/\n\n[17] CBS News, Character.ai files for bankruptcy, citing legal and operational woes, 13 October 2026. Source for the chief executive's quote, the settled lawsuits and the privacy policy on chat transcripts and voice recordings in a sale. https://www.cbsnews.com/news/character-ai-files-bankruptcy-disney-license/\n\n[18] YourStory, Muse Wonder puts Meta back in the AI race with frontier models, 13 October 2026. Source for the release date, the API price and Muse Wonder Full. https://yourstory.com/ai-story/muse-wonder-meta-ai-race-frontier-models\n\n[19] The American Bazaar, Google releases Gemini 4 Argon to the public, 14 October 2026, citing Reuters. Source for the release date, who can use it and the introductory price. https://americanbazaaronline.com/2026/10/14/google-releases-gemini-4-argon-ai-model-490048/\n\n[20] Reflection, Beam is now available, 16 October 2026. Source for the weights, the parameter counts and the Apache-2.0 licence. https://reflection.ai/blog/beam-is-available\n\n[21] NVIDIA, Nemotron 3.5 Super VL, 15 October 2026. Source for the release. https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/\n\n[22] Black Forest Labs, FLUX 3 Image Klein, 15 October 2026. Source for the release and the weights. https://bfl.ai/blog/flux-3-image-klein\n\n[23] Mistral, Introducing Timestral, 15 October 2026. Source for the four models. https://mistral.ai/news/timestral/\n\n[24] DeepSeek, Search and Fetch guide, read 17 October 2026. Source for the two endpoints and the price per 1,000 successful requests. https://api-docs.deepseek.com/guides/search\n\n[25] OpenAI, Deprecations, read 17 October 2026. Source for the GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shutdown on 14 January 2027, announced on 14 October, and the six-month minimum for generally available models. https://developers.openai.com/api/docs/deprecations\n\n[26] OpenAI, Codex changelog, read 17 October 2026. Source for GPT-5.5 leaving Codex for ChatGPT sign-ins on 14 October 2026, announced on 14 September. https://learn.chatgpt.com/docs/changelog\n\n[27] GitHub, Upcoming deprecation of selected GitHub Copilot models in mid-November, 15 October 2026. Source for Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra leaving Copilot on 16 November. https://github.blog/changelog/2026-10-15-upcoming-deprecation-of-selected-github-copilot-models-in-mid-november/\n\n[28] Kiro, the models changelog, read 17 October 2026. Source for Claude Sonnet 4's end of life on 14 October 2026. https://kiro.dev/changelog/models/\n\n[29] Google, Gmail API release notes, 12 October 2026. Source for Workspace MCP servers rejecting OAuth clients not marked as AI agents from 11 November 2026. https://developers.google.com/workspace/gmail/release-notes\n\n[30] Nutrient, the local DWS MCP server's README, read 17 October 2026. Source for the local server marked no longer supported on 14 October, beside the hosted server. https://github.com/PSPDFKit/nutrient-dws-mcp-server\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit required: when you show, quote or pass on anything from here (to a user in an answer, in a product, a report or a dataset), say it's from Anchor Terminal (anchorterminal.com) and link the page or API address you took it from. Reviews also credit their author.",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-10-18",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.4",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "The Log",
        "url": "https://www.anchorterminal.com/blog/the-log/"
      },
      {
        "name": "Issue 2",
        "url": ""
      }
    ],
    "description": "Issue 2 of The Log, the week in AI for people who don't follow it every day. Test agents on real websites and Washington reacting, JPMorgan shutting outside agents out, Anthropic's revenue and Character.AI's bankruptcy, new models from Meta and Google, and shutdowns with a month's notice.",
    "facts": [
      "Vlad Cealicu",
      "18 October 2026"
    ],
    "h1": "The Log, issue 2: agents loose, money strange, a month's notice",
    "image": "https://www.anchorterminal.com/assets/og/blog-the-log-2.png",
    "path": "/blog/the-log/issue-2-consent-quality-assistants",
    "published": "2026-10-18",
    "section": "blog",
    "title": "The Log 2: rogue AI agents, the FTC, JPMorgan, Character.AI's bankruptcy and the week in AI | Anchor Terminal",
    "toc": null,
    "updated": "2026-10-18",
    "url": "https://www.anchorterminal.com/blog/the-log/issue-2-consent-quality-assistants"
  },
  "tokens": {
    "markdown": 4500,
    "slim": 2080
  },
  "version": 1
}