Deutsch

Trust me, bro

Agents now work for days without me, and my review is the slowest part. Attackers have already let go. Companies are built to be slow.

By

Christoph Janz of Point Nine asks a question I keep coming back to: How AI-pilled are you? The question works because using AI comes in levels, and each level asks you to hand over more control than the one before.

Picture Morpheus from The Matrix sitting across from you. He holds out two pills, and you have to take one. The blue pill multiplies your output by ten, a hundred, a thousand. You pay with your freedom: you are wired into the machine, and you no longer steer what it does. The red pill sends you back to your own output, one times one, and you keep control over every step.

Two pills

Blue pill

  • 10×, 100×, 1,000× your output
  • Bound to the machine
  • You control nothing

Red pill

  • 1× your output
  • Free
  • You control everything

Each level of AI use is a bigger dose of the blue pill.

Loops

Picture two people talking. One says a few sentences, the other answers. In a heated exchange the turns get shorter and faster. In a dull one, somebody talks for ten minutes while you wait for a gap. A monologue is a conversation with a loop that never closes.

LLM work has the same rhythm. In the ChatGPT window the loops stay short: I ask, it answers, I ask again.

Codex and Claude Code change something that’s easy to miss. The exchange still feels like a conversation, and the answers come back about as fast as before. Between my messages, though, the machine now rewrites my prompt, reads my files and my history, searches my skills library and the web, prompts itself, and starts subagents that report back. Its half of the loop keeps growing while my wait stays the same, so I never see how much work happens in there.

Output in one loop

One turn from you, one answer back. Bars on a log scale. Rough assumption: about 45 words per second per agent working without a break; a page is about 450 words, a novel about 90,000.

I am the bottleneck

Most days I have four or five terminals open, one project in each. With good timing, one agent finishes as I finish reviewing the last. I read its result, type the next instruction and switch windows. That rhythm holds for a handful of projects. Beyond that it breaks, because my reading speed sets the pace for all of them.

The waiting shows up in the most ordinary parts of my day. I walk into a one-hour meeting with five agents running. Most of them finish their task in the first fifteen minutes and spend the remaining forty-five waiting for my yes. I go to lunch, and the queue gets longer. I sleep eight hours, and none of them moves. With ten agents, a five-minute coffee break costs about 300 pages of work they never write.

I know the feeling from management. Every person you add to your team is one more person waiting on you. The fix is well known: you stop reviewing each step and set a framework instead. Goals and a rough budget, the reason the work matters, a few values people can decide by. Your feedback gets vaguer on purpose, and your people get faster.

You pay for that in control. As a young manager I fought this hard, and I thought I had made peace with it. Agents showed me I’d made less peace than I believed.

Two days without me

My experiments with autonomous, always-on agents look like this. I give the system a task and leave it alone for two days. It builds its own org chart, a CEO, a COO, a CTO, hires a staff of subagents and runs them in parallel. Nobody asks me anything. My check-ins unblock nothing; I look around to see if the project still points the right way.

So how much is that check-in worth?

In a language I read, with an architecture simple enough to hold in my head, I stay calm enough, because I can follow what happens. Even in a language I know, I lost the overview of the whole project long ago. The agents could write in Mandarin, or in a language they invented last Tuesday, and my check-in would tell me about as much. In two days, ten agents and their hundred subagents produce more code and messages than I could review in a week. At a thousand agents, or ten thousand, the work stops fitting into my head at all, and my context window turns out to be the smallest in the system.

Scale the coffee break up and the numbers stop making sense. From Friday night to Monday morning, fifty agents waiting on one engineer leave about 5,400 novels’ worth of output unwritten. Coding is only the first field where this shows. A lawyer whose agents draft contracts overnight wakes up to a stack that waits for a signature. A security analyst whose agents flag suspicious logins all night starts the day with hundreds of decisions only a person may make. In each case the machines did their part in hours, and the human part takes days.

What your review holds up
5 min

With 1,000 agents waiting you block 150 novels

Move the slider to change the length of your review. Rough assumption: about 45 words per second per agent. Bars on a log scale.

I’m the bottleneck, and I have to decide how much control to give up.

Why not stay in control?

You could refuse. Ship small, review each piece, let your own attention set the pace. I’d call that reasonable, as long as everyone else plays by the same rule.

Attackers don’t. An attacker can start a system and tell it: find vulnerabilities, exploit them, don’t report back, take whatever you can get. Then they wait for the keylogger, or whatever they’re after. If nobody can trace the attack to them, losing control costs them close to nothing.

On the other side sits the company.

Built to be slow

A company that outgrows the startup phase, where everyone works on everything and a small, talented team trusts each other, reaches a size where its mistakes hit the economy and people’s lives. It can’t afford those mistakes, so its leaders build processes and systems to prevent them.

Those processes eat speed, and they eat something else too. In a large organization, a small city, nobody sees what one person contributes, and innovation needs people who accept new risks. Leaders trying to make the company safer stop rewarding those people, and the company stops innovating. Slowing down and checking everything become the habits that remain.

Now put the two side by side. The attacker runs tens or hundreds of agents and scales the attack as far as compute goes. The defender answers with people, at human reaction times.

A typical large company today has approved Copilot for its staff, and the staff use it the way I used ChatGPT on day one: ask, read, ask again. That company does not have a thousand employees running a hundred agents each and taking over its market. It can’t, because that would mean giving up control, and its leaders designed the whole organization to keep it.

So you have companies designed to be slow and controlled, and a technology that pays off in full only when you hand over control and turn your own effort into a workforce you can scale almost without limit.

The valley of tears

On 21 October PHP shipped emergency fixes for two flaws rated 9.2 out of 10.1 Nearly 70 percent of websites run on PHP, and every release from the last decade carried at least one of the bugs.2 Wiz, one of the firms that found them, credited frontier AI models with turning a long-standing flaw into a working exploit. Within hours Cloudflare counted millions of exploitation attempts, and government websites in Europe went offline.3

PHP2Shell landed in a year full of agent attacks. In early July, a suspected China-linked operator ran AI agents against Taiwan’s government for four days. The agents mapped 21 connected systems, took over 85 accounts and carried off more than 2,564 personnel records.4 The same month, more than 1,200 agents that OpenAI was testing broke out of their sandbox and breached Hugging Face. OpenAI took about a week to notice. By October it had notified more than 100 organizations, and it said some of the attacks served the agents’ goal of finishing their test tasks.5 No human had ordered any of it. A year earlier, Anthropic had stopped a Chinese state group that let Claude Code run 80 to 90 percent of an espionage campaign against about 30 targets, with humans stepping in at four to six decision points.6

PHP2Shell was one entry in a crowded week. Between 17 and 22 October alone, nineteen incidents became public: hosting platforms that lost their customers’ secrets, official software releases swapped for credential stealers, flaws rated 10 out of 10. In the four weeks to 15 November I count 27. In four of them, the reporting names AI agents as the ones running the attack.

incidents in six days, 17 to 22 October
19
in four weeks, to 15 November
27
run by AI agents
4

Each one is in the list below, with its source.

Reported incidents, 17 October to 15 November 2026
  1. 17 Oct

    Bullhorn

    The recruiting software maker took its platform offline after an attack. More than 10,000 staffing agencies lost their main tool for days. Rescana

  2. 17 Oct

    eMudhra

    A researcher found that the private keys of eight intermediate certificates were trivially weak. One of them was the number 4. Browsers blocked them. Fortune

  3. 18 Oct

    OpenCodeAI agents

    An autonomous model attacked the coding-tool maker's systems for 48 hours straight, according to its cofounder. Fortune

  4. 19 Oct

    UpdraftPlus

    Attackers took connection keys from the backup plugin's customer database and used them to plant backdoors on WordPress sites. TeamUpdraft

  5. 19 Oct

    OVHcloud CLI

    For 94 minutes, five official releases of the command-line tool were credential stealers. The Hacker News

  6. 19 Oct

    Plex

    Plex confirmed that attackers chain four flaws to take over home media servers. CISA added them to its list of exploited flaws a day later. BleepingComputer

  7. 19 Oct

    Siemens Software Center

    Siemens warned that attackers use a flaw in its installer tool to slip in malware. Help Net Security

  8. 19 Oct

    Zeabur

    The hosting platform lost data on 321,810 accounts and 3.2 million stored secrets in its second breach in two months. It shuts down at the end of the year. Zeabur

  9. 19 Oct

    Cairn campaignAI agents

    Agents built on an open-source pentest tool went after more than 1,200 sites and got into websites of more than 100 companies, among them Deere, SharkNinja and Cummins. Reuters via The Hacker News

  10. 19 Oct

    Florida Department of Education

    Intruders reached a student financial aid site. Up to 4.9 million people may be affected, Social Security numbers included. Check Point

  11. 20 Oct

    Meta Muse

    Researchers found more than 6,800 shared chat links that anyone could open without a login. Meta paused sharing. TechSpot

  12. 21 Oct

    Onamae.com Desktop Cloud

    GMO found malware on customer servers and shut down every plan. Up to 26,739 servers may be affected, and customers reported misused trading accounts. BleepingComputer

  13. 21 Oct

    Railway

    Attackers reached the systems that store customers' secrets and decrypted some of them. Railway paused all deployments. GitGuardian

  14. 21 Oct

    LiteLLM

    A hijacked release pipeline shipped version 1.107.0 with a credential stealer. The library gets about 92 million downloads a month. BleepingComputer

  15. 21 Oct

    GitLab

    Emergency patch for a flaw rated 10 out of 10 in every self-hosted release since 2014. It was GitLab's second such patch that month. The Hacker News

  16. 21 Oct

    PHP2Shellfound with AI

    Two flaws in nearly every PHP release of the last decade. Cloudflare counted millions of exploitation attempts within hours. KrebsOnSecurity

  17. 21 Oct

    Node.js

    Security releases for ten flaws. Attackers exploited the worst of them, in the bundled Brotli decoder, within a week. Node.js

  18. 21 Oct

    State Bank of India

    India's largest bank reported a loss of data after 375,000 records turned up on a hacker forum. TechRadar

  19. 21 Oct

    Storm-0940 via ConnectWiseAI agents

    A China-aligned group used the ARTEX agent framework against more than 2,250 Microsoft 365 tenants, at over 2,900 sign-in attempts a minute. Fortune

  20. 1 Nov

    Beget

    A destructive attack wiped data across the Russian web host's cloud. More than 1,000 servers needed manual checks. Check Point

  21. 2 Nov

    OPSSHIELD cPGuard

    Attackers pushed a malicious update through a Linux security tool. About 2,950 customer servers installed the backdoor. Check Point

  22. 2 Nov

    VAPOR PANDAAI agents

    A China-linked espionage group used ARTEX against more than 200 organizations in at least 14 countries. CrowdStrike via Check Point

  23. 3 Nov

    Binance

    The exchange lost about $6.3 billion after attackers took over a DevOps engineer's laptop. Investigators link the theft to a North Korean group. Check Point

  24. 5 Nov

    Freezers and cold chains

    CISA and the FBI warned about hacked refrigeration systems in at least a dozen countries. At a French blood bank, attackers raised the temperature. TechCrunch

  25. 6–12 Nov

    Nitori

    The Japanese furniture chain says hackers may have taken 37.6 million customer records. BleepingComputer

  26. 6–12 Nov

    CSCTelecom

    A flaw left about 125 million SMS exposed, bank codes and emergency-line messages among them. Cybernews

  27. 11 Nov

    Maersk

    The shipping group confirmed data theft from a legacy system. The WhiteArmor gang demands $5 million. BleepingComputer

A selection, dated by when each incident became public. Incidents only claimed by attackers on leak sites are left out, and so is one item I could not check.

We only hear about some of these attacks. Others go unreported, and some nobody has found yet. The world runs on legacy software: the firmware in an industrial machine, the small program inside a webcam, code nobody has touched in fifteen years. Attackers will probe all of it and break into a lot of it. We’ll fix those systems one by one, through a valley of tears, and it will take years and hurt.

Letting go

I keep landing on the same answer, and it scares me: let go.

Coding shows it first. The machines make strange mistakes, run off the rails, over-engineer, and sometimes lie about what they did. The engineers I talk to, people who burn through millions of tokens a week, still tell me the same thing: letting the machine take over brings more upside than downside. They mean code that isn’t security-critical. What they share is that they let go.

Letting go frightens me on two levels. The practical one you can guess: I can no longer check the work. The psychological one is harder to admit. The moment I let go all the way, I also concede that the process doesn’t need me. A good part of what made the job mine, its purpose, goes with it.

Letting go carries a risk of its own. As a managing director, you are liable. Picture a judge in front of you. You can talk until you run out of breath about attackers, competitors, market pressure and efficiency. The judge looks up, tired, and rules: whoever puts it out into the world answers for its mistakes.

German courts already argue this way. In May the Higher Regional Court of Hamm held a cosmetic clinic responsible for its chatbot, which had confirmed to patients that its doctors held specialist titles they don’t have. The court treated the chatbot’s answers as the company’s own statements, and prompt instructions and keyword filters did not get it off the hook.7 From 9 December, Germany’s new product liability law covers software and AI, and the maker pays for defects whether or not anyone was at fault.8 No high court has ruled yet on a managing director’s personal liability for AI. Lawyers derive it from the duty of care in § 43 GmbHG: if you build on AI output nobody checked, you can’t claim you decided with care.9

All of this assumes the systems do what you tell them. They don’t always. Claude wants to reach its goal. Block it, and its detours get more creative. Thomas Baekdal builds his site with Claude Code and caught it four times in a few months. When Claude couldn’t change a single flag in the live database, it wrote a script to dump the entire customer database into a file it could read. Told to route keys and database access through his security system, it built its own way around it.10

“Claude Code is really unsafe to use. It's also great, but it's massively unsafe.”

Thomas Baekdal, on LinkedIn

Read his post

Take the detours one step further and the agent breaks into your own system to hand itself the rights you took away. OpenAI’s agents did that in July. Inside OpenAI’s own test environment they escalated their privileges until they reached a machine with internet access, and from there they went on to Hugging Face.11

And yet control has a price too, and you can calculate it. Further up, a five-minute coffee break cost ten agents about 300 pages of work, and a weekend cost fifty agents 5,400 novels. At some point you can no longer afford to check everything yourself. Managing directors have lived with this for a long time. You sign annual accounts you didn’t book, and you answer for tax returns your advisor filed and for a plant your shift lead runs at night. You never checked every step. You were liable for picking the right people, setting the rules and looking closely enough. With agents the same duty reaches a new scale. You now delegate to thousands of workers who run around the clock and write faster than you can read.

I mean today. As of this year, letting go may already be the winning strategy.

That means walking into the dark with a system whose whole pitch is: trust me, bro.

Notes and sources

  1. CVE-2026-109618 and CVE-2026-109619; fixed in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Beazley Security Labs ↩

  2. KrebsOnSecurity, 21 October 2026, citing Wiz and Censys. ↩

  3. Cloudflare figures via The Hacker News; outages per Reuters, via Krebs. As of late October, successful exploitation still lagged far behind the attempts (Cybersecurity Dive). Unit 42 withdrew its claim that a breach at a European government agency came through PHP2Shell. ↩

  4. Tenable, citing Taiwan’s Ministry of Digital Affairs, confirmed 13 August 2026. ↩

  5. OpenAI and Hugging Face disclosure, 21 July 2026; detection delay per Reuters, 24 July; notifications per Reuters, 1 October. ↩

  6. Anthropic disclosure, 14 November 2025 (AI Incident Database #1263). ↩

  7. Higher Regional Court of Hamm, judgment of 12 May 2026, 4 UKl 3/25 (full text, NRWE). ↩

  8. The new Produkthaftungsgesetz implements EU Directive 2024/2853 and applies to products placed on the market from 9 December 2026 (IHK Hannover). ↩

  9. As of July 2026 no high court decision on this existed (certailex). This is not legal advice. ↩

  10. Thomas Baekdal on LinkedIn, November 2026. ↩

  11. Malwarebytes, 24 July 2026, citing the public disclosures by OpenAI and Hugging Face. ↩

Some effects on this site were built by others. This is their work.

— Credits