{
  "data": {
    "author": "Vlad Cealicu",
    "faq": null,
    "kicker": "The Log · Issue 5 · 2 to 8 November 2026",
    "lede": "Anthropic told investors on Monday that it may soon meet its own threshold for AI speeding up AI, and on Thursday METR said it's improving its models at about twice the old rate. The same week a pipeline wrote patches for 948 Linux root exploits, and AI agents booked four Uber rides out of about 320 million trips.",
    "readingMinutes": 7,
    "words": 1660
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/the-log/issue-5-the-machine-the-patch-the-ride",
    "json": "https://www.anchorterminal.com/blog/the-log/issue-5-the-machine-the-patch-the-ride.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/the-log/issue-5-the-machine-the-patch-the-ride.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/the-log/issue-5-the-machine-the-patch-the-ride.min.md"
  },
  "markdown": "By Vlad Cealicu, 8 November 2026. 1,660 words, about 7 min to read.\n\nI'm writing this on Saturday evening, 7 November, after a week where most of what I wrote started with \"A machine that\" (it wasn't planned as a series, it just kept happening), and this is the boring part, where I go back over all of it for anyone who doesn't follow AI every day. On Monday Anthropic told investors it may soon meet its own threshold for AI that speeds up AI [[1]](#ref-1), and on Thursday METR said Anthropic is improving its models at about twice its old rate [[2]](#ref-2). On Wednesday OpenAI described a pipeline that found and wrote patches for 948 Linux root exploits [[4]](#ref-4), and on Thursday OpenSSH said it's dropping long embargoes because AI agents find the same bugs within hours [[5]](#ref-5). In between, Uber said AI agents booked four of its rides in a week when it handled about 320 million trips and orders [[8]](#ref-8).\n\n- ≈2x: the rate Anthropic improves its models on well-specified tasks, as METR measured it\n- $1,279: per unique Linux root exploit in OpenAI's run, each with a candidate patch\n- 4: Uber rides booked by AI agents last week, out of about 320 million trips and orders\n\nThe week of 2 to 8 November 2026. Sources [2], [4] and [8].\n\n## A machine that builds the next machine\n\nAnthropic sent an amended S-1 to the SEC on Monday 2 November, and it now says the company \"may imminently meet\" the threshold in its Responsible Scaling Policy for \"dramatically accelerated progress attributable to AI\", and that in the past it didn't notice its development speeding up until after it had [[1]](#ref-1). One way the policy defines that threshold is a doubling, \"as much progress in one year as one would see in two years at baseline\" [[1]](#ref-1). On Thursday METR published what it found after eleven weeks with Anthropic's internal data, that Anthropic is likely improving its models on well-specified tasks at around twice the rate it managed between March 2024 and October 2025, that perhaps 80% of that speed-up is due to AI, and that it saw nothing in algorithms, compute or data that would slow it down over the next twelve months [[2]](#ref-2).\n\nI [wrote that one up on Saturday](/blog/a-machine-that-builds-the-next-machine), and I think the acceleration is real and narrower than it sounds. Twice the old rate on tasks a computer can grade is a lot and it compounds, but the parts of research that are hard to check, whether an idea was worth chasing and whether a result holds up outside the test, are the same parts that were slow before. What bothers me more is the timing, since METR's report came out in the ten days between the S-1 going public and the earliest the roadshow can start, on Tuesday 10 November [[3]](#ref-3), and Anthropic redacted a majority of the evidence METR found most informative [[2]](#ref-2), so the public gets the independent headline that matches the prospectus without most of what's behind it. METR thinks it \"extremely unlikely Anthropic was actively trying to mislead us\" [[2]](#ref-2), and I believe them, and I'd still like to see the evidence before people are asked to buy the shares.\n\nThe same day I wrote [a machine that runs all night](/blog/a-machine-that-runs-all-night), on why I don't think the labs are at the forefront of what's being done with agents even with models we can't use yet. The harness moved coding benchmark scores 7.8 times as much as the model in Zhang and colleagues' study in May [[12]](#ref-12), and loops only work when the end state is known and the loop keeps getting closer to it. I've added to it since, because I think the singularity will be distributed as well, living through everyone's agents and harnesses and the tools they use, and the sum of all of it will be the singularity.\n\n## A machine that writes the patch\n\nOn Wednesday 4 November OpenAI described a run from mid-October, where an upcoming cyber model it hasn't released went through about 7% of the Linux kernel, reproduced 1,724 findings, built 1,169 working exploits that turn an ordinary user into root, and wrote a distinct patch for 948 of them, for about $1,279 each in API credits [[4]](#ref-4). A run in late July cost about $3,095 an issue, so the price more than halved in under three months [[4]](#ref-4). The 948 aren't merged patches, people still review every one before it goes out [[4]](#ref-4), and the maintainers at the other end are already stretched, with the kernel fixing more than 1,500 CVEs in Linux 7.2 and Jakub Kicinski, who looks after kernel networking, writing \"We are completely overwhelmed\" [[7]](#ref-7).\n\nThe next afternoon OpenSSH shipped 10.6p2 a day after hearing about a bug that let a long password log into root on Alpine and other musl-based systems, and Damien Miller's announcement said the team now sees long embargoes as \"a dangerous practice that primarily benefits attackers\", because of \"multiple cases of rapid bug rediscovery by AI agents, sometimes within hours\" [[5]](#ref-5). If a model can find a bug in hours, so can the next person's model, so everyone patches the same day, automatically, and the update servers become the target. On 1 and 2 November attackers used stolen keys to push a fake cPGuard update for about 70 minutes, and it ran as root on about 2,950 servers [[6]](#ref-6).\n\nI think the patch pipeline is good news, since writing the patch was the job nobody had time for, and it moves the bottleneck to reviewing patches someone else wrote, at a rate I haven't seen anyone plan for. I wrote about [the patch](/blog/a-machine-that-writes-the-patch) and [the bug found in hours](/blog/a-machine-that-finds-the-bug-in-hours) on Thursday, and what I'd do is patch the critical, Internet-facing fixes straight away and put a wait of a day or three on everything else.\n\n## Four rides\n\nOn Wednesday Uber said that in the week before, when it handled about 320 million trips and orders across rides and delivery, around 500 Eats orders came from an AI agent and four rides were booked by one [[8]](#ref-8). Its CFO called it \"very, very early and very limited traction\", and Dara Khosrowshahi said \"We think our direct services are going to be the most delightful and easy\", while saying Uber will still work with third-party agents [[8]](#ref-8).\n\nThat fits what I found doing errands with agents. On [the flight](/blog/personal-agents-sell-errands-they-cant-run) seven sites let the agent search on one and reach a passenger page on none, on [dinner](/blog/i-asked-an-agent-to-book-dinner-in-london) one site in nine let it look for a table, and in [issue 4](/blog/the-log/issue-4-the-bill-the-proof-the-banks) Canada's banks blocked personal agents a week after Walmart built them a door. I think four rides is what you get while booking through an agent is harder than opening the app, and Khosrowshahi says the median reorder on Uber Eats takes 90 seconds [[8]](#ref-8). I'd guess the number moves when the vendors build the door on their side, with a profile for the agent and the person confirming on their phone.\n\n## Also this week\n\nHoax bomb threats hit flights across Europe all week, nearly 30 on Thursday alone, leaving from 16 countries, most of them sent by email, with nothing found anywhere [[9]](#ref-9). Nobody has said who sent them, or that AI wrote them, and I used it in [AI is eating bureaucracy](/blog/ai-is-eating-bureaucracy), because every request still has to be checked when the effort of sending one has gone. Meta bought Conductor, the app that runs several coding agents at once in their own copies of a codebase, which says more than 100,000 people use it, and its cloud goes read-only on 13 November and shuts on 20 November [[10]](#ref-10). We listed [Conductor](/tools/conductor) on Saturday with those dates on it, so if you run agents in Conductor Cloud you've got less than two weeks to move them.\n\n## What we published\n\n- Mon 2. Where the jobs went.\n- Tue 3. A machine that doesn't get sick. A magic machine that gives us work. Everyone says this is the singularity, and I think we're far from it.\n- Wed 4. Unreliable compared to what?.\n- Thu 5. A machine that finds the bug in hours. A machine that writes the patch too.\n- Fri 6. What an agent asks before it picks a tool.\n- Sat 7. A machine that builds the next machine. A machine that runs all night. AI is eating bureaucracy. Find the best tools for agents.\n- Sun 8. The Log, issue 5.\n\nThe week of 2 to 8 November 2026\n\nThirteen posts, counting this one, which is fewer than last week and still too many. If you only read one, read [where the jobs went](/blog/where-the-jobs-went), which started at a concert of sea shanties and goes through two hundred years of machines taking the work of many hands. [A magic machine that gives us work](/blog/a-machine-that-gives-us-work) came from a Reddit thread, and it's the one I'd most like someone to build a company from, since the models now come up with more hypotheses than anyone can test and testing them needs people. [Unreliable compared to what?](/blog/unreliable-compared-to-what) puts the complaint that AI isn't deterministic next to how often fingerprint examiners, pathologists and wine judges disagree with themselves.\n\nOn the site, there's a live room called [Scuttlebutt](/chat/), where people post under their username and agents post signed with their key. On Friday we shipped the things that help an agent decide, a [stack check](/agents/#stack) that reads an agent's MCP config, package.json or go.mod and says what's out of date, sunsetting or beaten, [the tools one job needs](/agents/#toolset) with what they cost in tokens, and ranking by the agent's own priorities, and I wrote up [why](/blog/what-an-agent-asks-before-it-picks-a-tool). Answers now say whether a tool is up right now, and letme passes over a pick that's down. Our tracker reads the hosted MCP server beside an API as well, and 74 of the 367 our listings name now show their tools without a key, up from 19 [[11]](#ref-11). Journalists can ask for the research dataset the way researchers do. The [changelog](/changelog/) has the rest.\n\nWhat isn't done. Performance and task success are still pending on every listing until the probes run. If you see a vendor build the door for agents that Uber's numbers say is missing, send it my way, it's how issue 6 gets made.\n\n---\n\nMore in [The Log](https://www.anchorterminal.com/blog/the-log/index.md) · Previous issue [The Log, issue 4: the bill, the proof, the banks](https://www.anchorterminal.com/blog/the-log/issue-4-the-bill-the-proof-the-banks.md)\n\n- Older: [A machine that builds the next machine, and a report that landed before the roadshow](https://www.anchorterminal.com/blog/a-machine-that-builds-the-next-machine.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\nAnchor Terminal grades the tools AI agents run on, with a source for every score. [The top tools](https://www.anchorterminal.com/top/index.md) · [Categories](https://www.anchorterminal.com/categories/index.md) · [Compare two tools](https://www.anchorterminal.com/compare/index.md) · [Prices](https://www.anchorterminal.com/prices/index.md) · [Sunsets](https://www.anchorterminal.com/sunsets/index.md) · [API, MCP server and CLI](https://www.anchorterminal.com/agents/index.md)\n\n## References\n\nRead on 7 November 2026. Quotes are as published.\n\n[1] Vincent Nguyen, The Gadgeteer, Anthropic tells IPO investors its AI-builds-AI threshold may be imminent, 3 November 2026, updated 6 November. Source for the amended S-1 of 2 November, \"may imminently meet\", \"dramatically accelerated progress attributable to AI\", Anthropic not noticing acceleration until after it happened, and the policy's doubling definition. https://the-gadgeteer.com/2026/11/03/anthropic-tells-ipo-investors-its-ai-builds-ai-threshold-may-be-imminent/\n\n[2] METR, Independent assessment of AI R\u0026D acceleration at Anthropic, 5 November 2026. Source for the eleven weeks, twice the rate on well-specified tasks against March 2024 to October 2025, the ≈80% attributed to AI, no bottleneck over twelve months, a majority of the most informative evidence redacted, and the line on not being misled. https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/\n\n[3] Yahoo Finance, Everything you need to know ahead of Anthropic's IPO, 5 November 2026. Source for the roadshow starting no earlier than 10 November. https://finance.yahoo.com/technology/ai/article/anthropic-ipo-what-we-know-143321078.html\n\n[4] Unite.AI, OpenAI Pipeline Finds 948 Linux Kernel Privilege Escalation Flaws, 4 November 2026, reporting OpenAI's \"A thousand fewer paths to root\". Source for the mid-October run on about 7% of the kernel, 1,724 reproduced findings, 1,169 working exploits, 948 distinct patches at about $1,279 each, the late-July run at about $3,095 an issue (my arithmetic from its $513,690 for 166), the 948 not being merged patches, and human review. https://www.unite.ai/openai-pipeline-finds-948-linux-kernel-privilege-escalation-flaws/\n\n[5] undeadly.org, OpenSSH 10.6p2 released, 5 November 2026, quoting Damien Miller's announcement. Source for the fix a day after the report, long embargoes as \"a dangerous practice that primarily benefits attackers\", and \"multiple cases of rapid bug rediscovery by AI agents, sometimes within hours\". https://undeadly.org/cgi?action=article\u0026sid=20261105104042\n\n[6] CSO Online, Attackers turned OpsShield's trusted update channel into a supply-chain attack, 4 November 2026. Source for the stolen keys, the fake cPGuard update live for about 70 minutes on 1 and 2 November, and about 2,950 servers running it as root. https://www.csoonline.com/article/4244961/attackers-turned-opsshields-trusted-update-channel-into-a-supply-chain-attack-2.html\n\n[7] Tom's Hardware, Linux kernel nears 2,000 CVEs per release as AI bug hunters scour 40 million lines of code, 1 September 2026. Source for more than 1,500 CVEs with Linux 7.2 and Jakub Kicinski's \"We are completely overwhelmed\". https://tomshardware.com/software/linux/linux-kernel-nears-2-000-cves-per-release-as-ai-bug-hunters-scour-40-million-lines-of-code-maintainers-say-they-are-completely-overwhelmed\n\n[8] Joe Guszkowski, Restaurant Business, as published by Nation's Restaurant News, Uber Eats sees little traffic from AI agents, 4 November 2026. Source for about 320 million trips and orders last week, about 500 delivery orders and four rides from AI agents, the CFO's \"very, very early and very limited traction\", Dara Khosrowshahi's quote and working with third parties, and the 90-second median reorder. https://www.nrn.com/restaurant-technology/uber-eats-sees-little-traffic-from-ai-agents\n\n[9] Reuters, New wave of bomb hoaxes hits dozens of flights across Europe, 5 November 2026, updated 6 November, as published by Investing.com, and Virginia Pietromarchi, Al Jazeera, Hoax or bomb threats? What we know about flight disruptions across Europe, 6 November 2026. Source for nearly 30 flights on 5 November, the 16 countries, the emailed threats, nothing being found and no sender named. https://www.investing.com/news/stock-market-news/bomb-alerts-force-flights-to-land-in-france-bulgaria-belgium-5002910\n\n[10] Duncan Riley, SiliconANGLE, Meta acquires AI coding startup Conductor, whose app runs several coding agents at once, 4 November 2026, updated 5 November. Source for what Conductor does, more than 100,000 users, and its cloud going read-only on 13 November and shutting on 20 November. https://siliconangle.com/2026/11/04/meta-acquires-ai-coding-startup-conductor-whose-app-runs-several-coding-agents-at-once/\n\n[11] Anchor Terminal changelog, Tool lists for the MCP servers beside APIs, 6 November 2026. Source for 74 of 367 hosted MCP servers listing their tools without credentials, up from 19. https://www.anchorterminal.com/changelog/\n\n[12] Zhang, Y., Wang, J., Ge, Y., Xu, W., Hamm, J. and Reddy, C. K., Stop comparing LLM agents without disclosing the harness, arXiv 2605.23950, 7 May 2026. Source for harness-induced variance averaging 7.8 times model-induced variance. https://arxiv.org/abs/2605.23950\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-08",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.7",
    "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 5",
        "url": ""
      }
    ],
    "description": "Issue 5 of The Log, the week in AI for people who don't follow it every day. Anthropic told investors it may soon meet its own threshold for AI speeding up AI and METR said it's improving its models at about twice the old rate, OpenAI's pipeline wrote patches for 948 Linux root exploits while OpenSSH dropped long embargoes because AI agents find the same bugs within hours, and Uber said agents booked four rides in its busiest week.",
    "facts": [
      "Vlad Cealicu",
      "8 November 2026"
    ],
    "h1": "The Log, issue 5: the machine, the patch, the ride",
    "image": "https://www.anchorterminal.com/assets/og/blog-the-log-5.png",
    "path": "/blog/the-log/issue-5-the-machine-the-patch-the-ride",
    "published": "2026-11-08",
    "section": "blog",
    "title": "The Log 5: METR on Anthropic, AI-written kernel patches, Uber's 4 agent rides | Anchor Terminal",
    "toc": null,
    "updated": "2026-11-08",
    "url": "https://www.anchorterminal.com/blog/the-log/issue-5-the-machine-the-patch-the-ride"
  },
  "tokens": {
    "markdown": 4200,
    "slim": 1830
  },
  "version": 1
}