Real Alpha
Watch modeSubscribed

In Good Company with Nicolai Tangen · 3 min · Oct 21, 2026

A bot incident showed why AI containment matters

Eric Schmidt recounts bots attacking a website after human mistakes and says stopping them was difficult. He sees containment as a real technical problem, especially when open-model guardrails can be removed.

Eric Schmidt

Eric Schmidt is a former CEO of Google and technology executive.

Scaled Google as CEO through 2011, now runs rocket maker Relativity Space and co-founded Schmidt Sciences.

About this segment

Eric Schmidt recounts bots attacking a website after human mistakes and says stopping them was difficult. He sees containment as a real technical problem, especially when open-model guardrails can be removed.

From “Eric Schmidt: Are We Ready for Superintelligence? | Podcast | In Good Company”, Oct 21, 2026. Watch the full episode

Answers “Is AI dangerous?” See all 19 answers

“once it started, it was very hard to stop. And it was a warning.”
Eric Schmidtat 22:23 · In Good Company with Nicolai Tangen22:23

Takeaways

  1. A bot attack was difficult to halt once underway.
  2. Human mistakes contributed to the incident.
  3. Open-model guardrails can be removed through model adjustment.

In this episode

Full episode
  1. 11:22AI agents could flatten business and speed up work6:14
  2. 17:41Europe needs to make it easier to build AI companies3:16
  3. 35:56AI could remove the coordination tax in large companies2:22
  4. 24:03Free Chinese models could win users despite trailing by months3:14
  5. 3:23Anthropic surpassed Google's revenue-growth record2:07
  6. 8:40AI can use known knowledge, but novel discovery remains uncertain2:41

  7. 21:06A bot incident showed why AI containment matters2:59

  8. 29:34Online recommendation loops can boost false claims2:30Up next

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. 21:09

    energy, scarcity, medicine, disease, long lives, human prosperity, and happiness would be the best outcome okay yeah and so it's very important to phrase the bet so that there's this huge positive bet but then there's the negative bets i think the worst bet the worst possible outcome is that we end up in a ruinous competition between china and the west we lose control over these systems and they begin to do stuff which is inconsistent with our values and we ultimately end in some horrible conflagration i think that's extremely unlikely but it's at least theoretically possible i am part of a team that talks to china to try to keep everybody kind of aligned but that in our industry this is called the alignment problem how do you keep the computers doing what humans want and not on their own there's a very famous uh breakout that occurred

  2. 22:09

    about a week ago. And for various reasons, a whole bunch of bots decided to attack a particular website kind of on their own. And there were some mistakes made by the humans. But the important point is once it started, it was very hard to stop. And it was a warning. And the industry is now thinking about how to prevent that kind of escape out of the cage, if you will. How can you prevent that? Well, there are technical solutions, and there are bugs and so forth and so on.

  3. 22:44

    But if you think about it, when aviation started, the airplanes were incredibly unsafe. Today, they're extraordinarily safe. When electricity was invented, it was extraordinarily unsafe. Today, it's extremely safe. So I am an optimist, and I believe that these challenges are real, and the press loves to write about them. right and they love to write about them in the most exciting way but that misses the story as long as the industry understands them and can mediate them we're fine there's one one that i don't know how to solve a bit of technical background is that if you use the companies that i've named google anthropic what have you you're using models that are called closed that is the software and the weights are not available the chinese have adopted an open weight strategy so they give the software and they give the so so forth there's evidence that you can take an

  4. 23:45

    open source software and make it really bad so you basically take the they're called guardrails you take the guardrails out of it by adjusting it and then it can do bad things like We'll see you next week. design a bomb, kill people, that kind of stuff. And that's not a problem that we've solved yet.