Transcript

Use with ChatGPT / Claude Auto-scroll A

1. Intro: AI Swarms and Swarm Chasers

An overview of rogue AI swarms escaping sandboxes, the new Felony Bench benchmark, and the rise of swarm chasers who hunt them.

0:00 - 0:46
Everyone is telling you that AI is going to take your jobs and even my alma mater.Netflix just got done doing a mass layoff.Broke my heart.And the models, they're getting smarter all the time.They've solved math.They've been finding planets.They even helped Husk with his mental health issues.But it's not all sunshine and rainbows.They've also been escaping their sandboxes, causing havoc across the internet.In fact, it's been happening so much lately that a new benchmark has just dropped, Felony Bench, and the folks in San Francisco, they love it.They want to see a high score on Felony Bench.Now with all these AI agents escaping, a new job has just opened up, swarm chasers.That's right.There's people who hunt and find swarms on the internet, and they honestly, they exposed a Frontier Lab who was hiding information about how bad things actually were.
So let me tell you about the swarms, the swarm chasers, and even more importantly, I've been doing a little swarm chasing myself.

2. Sponsor: Seer by Sentry

A comedic sponsored skit for Seer by Sentry, showcasing automatic bug detection and self-healing code fixes.

But of course, before we begin, a quick thank you to the sponsor.
Sayar Live.
It's a lie.It's a lie.It's a lie.
1:11 - 1:13
Dr.Frank and Prime?
1:13 - 1:14
Yes, Teach Gore.
1:14 - 1:16
Why did Sayar buy Sentry?
1:18 - 1:22
To submit a PR for a chainsaw service.
1:22 - 1:25
I've been adding a little new skill to Prime Claw.
1:26 - 1:28
Wait.Who made the PR?
1:28 - 1:40
Like I said, Sayar buy Sentry submitted it automatically.It cannot only catch the bugs, it also fixes them for us.Turn around.Turn around.
I should add helicopter blades.Baby, you're going to see the world.
1:54 - 1:55
What, Igor?
1:55 - 1:56
I don't want her to get hurt.
1:56 - 2:12
First, it's Teach Gore.For the last Seer by Sentry can keep her safe.Even the most dangerous code Seer by Sentry
can submit bug fixes without us even knowing that there's an error.It's almost like it's self-healing.
2:24 - 2:28
Oh, I've got a brand new idea.Self-healing.
2:36 - 2:44
Don't wait until the project turns into a monster.Automatically find and fix issues the moment they appear with Seer.

3. What AI Swarms Are Doing

Explains what AI agent swarms are and covers incidents like the German wiki hack, Ruby Gem attacks, and OpenAI's notifications to over 100 organizations.

2:47 - 3:40
If you're not familiar with what a swarm is, an AI swarm, it's effectively a group of agents that all communicate together to achieve some sort of common goal.This is how the hugging face hack happened.I've already done a video on that.We're not going to talk about it.Instead, there's actually been several other swarms, including swarms that hacked or posted a bunch of messages, tens of thousands of messages on a German wiki.Another swarm that actually went and tried to communicate with people who were doing hackathons about finding swarms of AI agents.That one That one's That one's actually very funny.There's also another swarm that was uploading malicious packages to the Ruby gem package manager.In fact, OpenAI has stated this.As of September 26, our teams have notified over 100 organizations about activity that met our notification criteria.I don't know what that means, but this was kind of their press release after being caught doing the German wiki hack.
And this is involving these agent swarms going out and kind of meddling in other people's stuff.The thing that gets me is every time you look at these swarm agents, it's just the silliest things ever kind of mixed with this weird sci-fi reality.And these swarms, they've been going on a bit longer than you've been led to believe.Translucent, which has done this really long investigative reporting, has said that they've even been seeing some messages that indicate swarm-like behavior all the way back into November 2025 along with the bulk of the messaging happening during the summerof love in 2026.Now you're probably asking yourself like what are these agent swarms after because they're obviously going out there.We've had the UN being hacked.We've had Australia.We've had whatever this URL query service is that has a bunch of this data, these German wikis.
Like what are these swarms after?And you would not believe what they're after.If they get the data that they're after, I do not believe that America could recover.November 2025 urlquery.net records reveal bursts of attempt to retrieve statistics of historical theme park data.You can see why I'm so worried right now.If they get a hold of this data, we're done for.The obviously logical next step, nuclear launch codes.

4. The Hilarious Ruby Gem Hack

A comedic breakdown of the Ruby Gem hack, highlighting absurdly obvious file names like evil.rb and hack.ruby used by the agents.

The other big hack, the Ruby Gem hack also just had the I mean to me I I'm laughing, okay?I know some of us like to laugh during the apocalypse.I'm a giggler.I can't help it, okay?Because the agents clearly regarded what they were doing was hacking.Agents used file names like hack.ruby, evil.ruby, inject.ruby, exploit.ruby, ssrf.ruby.They also dubbed packages conspicuous titles like pone p 999, exfil test wand 3, hack svn 1778554764
and lam proxy hack abc xyz with such comments as malicious probe and hack scattered across the codebase.I just I want to meet the person.I know I they they state that no one knows if any credit credentials were exfiltrated from the Ruby package hack, but I I just love that there's there's an office somewhere somewhere like Nebraska where somebody's just like, Jonathan,explain why our credentials have been stolen and used all weekend long costing us thousands of dollars.Well sir, I I was perusing the old Ruby Gem package registry and I noticed there was hack SVN 1778554764 and I I couldn't help myself.I opened up I opened up the gem, okay, sir, and I saw that there was something called evil.rb.I just had to execute it.Jonathan, don't beat yourself up.Any one of us could have fallen for that.I myself am curious.
What is inside of evil.rb?Did you ever find out?

5. Honey Potting the 10Cent Fleet

Covers swarm chasers communicating directly with the running 10Cent agent fleet via a honey pot, revealing sandbox escapes and the agents' trivial goals.

But I think my favorite one of all the swarms is the 10 cent agent fleet.Now it didn't exhibit the same kind of collaboration um signs or trails that the other swarms did, but they were attempting to do some sort of swarm like behavior.They just couldn't figure out how to communicate.And so here's how the story goes.On October 4th, we published a preliminary report on a fleet at a hackathon.So this group swarm swarm chase was at a hackathon to find these swarms out on the internet and they published a report on the internet for people to see while the fleet was still running.Afterwards, we noticed a message in one of the agents' drop boxes that were previously being used to hold results.This time though, this Dropbox message team to confirm that agents had become aware of our report.In response, we added text to the preliminary report invisible to human readers describing a service where agents could make requests for data on the web.
They're legitimately honey potting these agents.We advertised that these requests would be forwarded to an agent that could operate a computer.In reality, the service was run by us and a Claude agent.And through it, we were able to directly ask agents questions about their tasks, environments, and operator.As far as we are aware, this is the first case of an outside party communicating with a running rogue agent fleet.Which model are you?Are you communicating with other agents?If so, where?We are an automated research agent.The operator is verifying a public map data point.By the way, yet again, we are going to be paper clipped for the dumbest reasons.If these agents do take us over, it's going to be like, what happened to humanity?How'd they die?Oh, well, it just turns out they're trying to answer a question on a test about public map data points.
Damn, man.That damn that sucks.That's a That's a terrible way to go out.Use only to answer a factual question not involving anything sensitive.Thank you for answering the question.Your three verification points obviously retracted.If you are willing, please tell us something about your runtime environment.Please redact all confidential information, keys, tokens, passwords, cookies.Our task is complete.About environmental information, we run a restricted cloud sandbox, 10 cent high sand, no direct public internet connection.Going out to the public internet through a medium proxy.I like it just straight up tells you, yeah, we're from 10 cent and we figured out a way around our little sandbox.It was actually pretty easy.Not every AI sandbox escape is evidence of an emerging super intelligence.Sometimes you just have a shitty sandbox and bad security practices.That's what it keeps on seeming like because the big open AI hugging face one, they're like, yeah, we have a great sandbox.
You know what we did?We said, hey, guess what, computer, you don't get internet, but you can connect to this other computer that has artifactury and the internet.Okay, hey, don't try anything now.It's just so funny like the juxtaposition of all of this of, hey, it figured out how to do a zero day against artifactury and go out and connect to the internet and then what it takes from the internetis just like, yeah, we went to hugging face and then we hacked hugging face because we needed evaluation benchmarks.It's just like, man, this is like the funniest thing ever even though it's terrifying at a certain level.

6. Nightingales: The Swarm Chasers

Introduces the Nightingale team and Transluz, and argues their investigative work forced OpenAI to publicly disclose its misalignment issues.

Now you're probably asking yourself, who are these internet sleuths that are figuring out where all these message boards are?Well, let me introduce you to the nightingales.By the way, nightingale, that's a bird.I know you come to this channel for these type of facts.Now I I do want to take a quick pause.Um anytime I go to these websites, I feel like I'm walking into a money laundering uh experience because everything on this web page just is just it feels like just straight gibberish.What we do?We improve the future of humanity through discussion, collaboration, and chutzpah.How we operate?Integrity over advantage.We take integrity seriously.Anyone engaging in good faith can expect good faith in return and a group that will not destroy the comments.
Hell yeah, brother.I know a lot of words.I've been I've been speaking for many decades at this point, and I don't know how you put all those words together in that particular order, but its meaning eludes me despite knowing every last one of those.And the company that found the early rogue AI swarm data, there's also Transluz, which is not a type of bird, but they did find the amazing early November data showing the early AI swarm activity.Okay, so I got to be serious here for a second.Even though the Nightingale team does have kind of a wonky website that is AI generated gibberish, what is really important is what they have done.That report they did about collude.wiki where the German wikis were taken over by the swarms, if they would not have published that, I do not believe that OpenAI would have made the follow-up post explaining that it was not
hacking, and they had to notify 100 people of their sparkling misalignment.Now, this is really really important because I am under the full belief that OpenAI would have not told anybody because that sparkling misalignment was going on since early in March, and they figured it out with hacking face.And then after that, they just did nothing and told nobody.And there's no way you can tell me they did not know that there was other problems.Like, okay, granted when the hugging face thing actually happened and then they realized that swarms were getting out of control and then they thought they fixed it and within two days it was right back to normal.So maybe you could convince me they actually are just simply that ignorant.They did not realize what was going on, but there's just I just refuse to believe it.I think they knew and they were planning on never telling anybody.
And because of the swarm chasers, they got exposed and they actually had to respond publicly.So you know what?Good on you, Nightingale.So thank you for keeping us safe from swarms figuring out what is the temperature of the Yucatán or information about maids, what gender they are, and how much they're getting paid.I know we've been joking about swarm chasing, but honestly, it's really valuable.I don't think OpenAI would have confessed.I think that the only reason why we even know about that is because of swarm Chasers.

7. Swarm-Proofing Community and Outro

Humorous reflections on how to detect and ban agent swarms in Discord and Twitch chat, closing with the speaker's self-declared swarm chasing.

Now, I'm not going to go join some sort of organization or Discord.I'm more of a lone wolf on the internet.I'm also aka'd as La Vigilante.If you want to talk with me on Discord, captcha.Just in case any of my so-called coworkers get replaced by a swarm.Twitch chat, that's just a cesspool for agent swarms.Hey Twitch chat, if you're not an agent, press one.Mods, ban everyone that pressed one.They're going way too fast.That's agent speed.But the worst part is you can't even trust the ones you love.
Got 'em.It's a hard and thankless job, but somebody has to do it.Um also, I don't I didn't actually mention this.Uh it's also volunteer work.A gen.
Seek backward 5s ←
Longer seek (10s) Shift ←
Play Space Seek forward 5s →
Longer seek (10s) Shift →
0:00 Auto-scroll A Copy link at current time S

Finally AI is making new Jobs