00:00:00A Different Kind Of Data Breach
00:01:07Meet Mercor And RLHF Power
00:03:25Lite LLM Becomes The Backdoor
00:05:44What Was Stolen And Who’s Hurt
00:07:50The Allegedly Fake Compliance Problem
00:09:39What You Can Do Now
درباره این اپیزود
A normal data breach steals names and passwords. This one may have stolen the recipe for building the world’s most powerful AI models, and it happened through software most people will never notice until it breaks. We follow the Mercor breach from the first warning signs to the moment poisoned Python packages hit PyPI and spread in minutes across systems that were set to auto-update.
We walk through what Mercor actually does in the AI economy, especially RLHF (Reinforcement Learning from Human Feedback), and why that behind-the-scenes work shapes how tools from OpenAI, Anthropic, Meta, and Google behave. Then we unpack Lite LLM, the open source “plumbing” that connects apps to multiple AI services, and how a supply chain attack can bypass the company you’re targeting by compromising the dependencies everyone trusts.
From there, the focus shifts to the fallout: contractors whose Social Security numbers and identity documents may be exposed, companies scrambling to assess backdoors and credential theft, and the bigger fear that proprietary AI training data sets and labeling strategies are being auctioned on the dark web. We also dig into the compliance controversy around SOC2 and ISO 27001 style certifications and what happens when security audits become performance instead of protection.
If you care about cybersecurity, data privacy, AI governance, and open source risk, listen through to the end for concrete steps you can take right now. Subscribe, share this with a friend who uses AI tools, and leave a review with your take on who should be held accountable.
انگلیسی
ایالات متحده آمریکا
یادداشت ها را نشان دهید 🔗
رونوشت 🔗
WEBVTT
00:00:00.239 --> 00:00:03.359
I'm pretty sure you've heard about data breaches before.
00:00:03.759 --> 00:00:04.160
Right?
00:00:04.639 --> 00:00:08.160
A hospital gets hacked, a retailer exposes credit card numbers.
00:00:08.560 --> 00:00:12.080
Someone's social security number ends up on the dark web.
00:00:13.119 --> 00:00:15.359
This one is definitely different though.
00:00:16.079 --> 00:00:20.239
So three weeks ago, hackers didn't just steal personal data.
00:00:20.719 --> 00:00:27.120
They may have walked out with the actual blueprints for how the world's most powerful AI models are built.
00:00:27.920 --> 00:00:33.119
The secret sauce behind open AI, behind Meta, behind Anthropic.
00:00:33.600 --> 00:00:34.000
Yeah.
00:00:34.399 --> 00:00:35.200
Those guys.
00:00:35.920 --> 00:00:43.840
And they got in through a piece of software almost nobody has ever heard of that was downloaded 97 million times last month alone.
00:00:44.719 --> 00:00:47.039
This is the story of the Mercor breach.
00:00:47.520 --> 00:00:52.000
And if you've been listening to this show, you're in it.
00:00:52.479 --> 00:00:53.520
Let's go.
00:01:07.120 --> 00:01:13.280
Alrighty then, ladies and gentlemen, welcome back to another episode of Privacy Please.
00:01:13.439 --> 00:01:18.640
I am your host, Cameron Ivey, and today we're following the money, baby.
00:01:18.879 --> 00:01:19.840
The data.
00:01:20.480 --> 00:01:28.959
And a hacking group that just pulled off what might be the most consequential AI heist in American history or world history.
00:01:30.159 --> 00:01:31.359
Let's go ahead and dig in.
00:01:31.439 --> 00:01:39.200
But before we do, I must remind you the Privacy Please podcast is part of the Problem Lounge Network.
00:01:39.439 --> 00:01:41.359
So if you're not familiar with it, check it out.
00:01:41.519 --> 00:01:49.599
The problemlounge.com, our website for our show, The Problem Lounge, and for Privacy Please, the show you're listening to.
00:01:49.920 --> 00:01:51.120
Go give it some love.
00:01:51.519 --> 00:01:52.560
Go follow us.
00:01:52.799 --> 00:01:55.519
Go find our YouTube, the Problem Lounge Network.
00:01:55.840 --> 00:01:56.640
We're on there.
00:01:56.959 --> 00:01:59.120
Alright, let's jump in.
00:01:59.680 --> 00:02:02.959
Most people have never heard of Mercor.
00:02:03.359 --> 00:02:05.200
That's kind of the point.
00:02:05.439 --> 00:02:10.639
They were founded in 2023 by three friends who met on their high school debate team.
00:02:10.800 --> 00:02:13.120
They were 19 years old when they first started it.
00:02:13.360 --> 00:02:19.919
Mercor sits at one of the most critical and least visible junctions in the entire AI economy.
00:02:20.319 --> 00:02:28.639
Here's what it does: it recruits experts, doctors, lawyers, engineers, scientists, writers, and pays them to help train AI models.
00:02:28.879 --> 00:02:36.400
They read outputs, rank responses, flag errors, provide high-quality human feedback that makes the AI smarter.
00:02:36.639 --> 00:02:44.080
That process is called RLHF, Reinforcement Learning from Human Feedback.
00:02:44.319 --> 00:02:49.199
And it's the engine behind the every major AI system you've ever used.
00:02:49.520 --> 00:02:54.400
Mercur's clients, OpenAI, Anthropic, Meta, and Google.
00:02:54.639 --> 00:02:56.400
Those are pretty big names.
00:02:56.639 --> 00:03:01.199
By October 2025, the company was valued at$10 billion.
00:03:01.840 --> 00:03:07.840
Its three founders, all 22 years old, became the youngest self-made billionaires in the world.
00:03:08.080 --> 00:03:09.520
And nobody's ever heard of them.
00:03:09.680 --> 00:03:20.479
And on March 27, 2026, it became the center of one of the most consequential cybersecurity incidents in the history of artificial intelligence.
00:03:20.879 --> 00:03:25.520
Here's where it gets fascinating and alarming in equal measure.
00:03:25.759 --> 00:03:28.879
The hackers didn't attack Mercor directly.
00:03:29.039 --> 00:03:34.719
They didn't break down the front door, they poisoned the water supply, so to speak.
00:03:34.960 --> 00:03:38.639
Their target was a piece of software called Lite LLM.
00:03:38.800 --> 00:03:41.919
Most people outside of software development have never heard of it.
00:03:42.080 --> 00:03:48.719
But if you've used any AI tool in the last two years, there's a very good chance it was running underneath it.
00:03:49.039 --> 00:04:00.800
Lite LLM is an open source Python library that lets developers connect their applications to AI services, OpenAI, Anthropic, Google, all of them.
00:04:01.360 --> 00:04:03.680
Through a single unified interface.
00:04:03.840 --> 00:04:07.280
It's downloaded 97 million times a month.
00:04:07.520 --> 00:04:11.840
It's estimated to be present in 36% of all cloud environments.
00:04:12.000 --> 00:04:15.680
It is essentially the plumbing of the AI industry.
00:04:15.919 --> 00:04:21.759
A hacking group called Team PCP had been working their way upstream for weeks.
00:04:21.920 --> 00:04:32.240
They first compromised a tool called Trivi, a security scanner, and used that to steal the credentials of a maintainer at Lite LLM.
00:04:32.560 --> 00:04:49.680
Then on March 27th, they used those stolen credentials to publish two poisoned versions of Lite LLM directly to PyPy, which is PYPI, the main repository where developers download Python software.
00:04:49.839 --> 00:04:53.839
The malicious code was live for approximately 40 minutes.
00:04:54.079 --> 00:05:07.680
Now, in those 40 minutes, because thousands of systems are configured to automatically pull the latest version of software they depend on, the poisoned packages were downloaded and installed across thousands of companies.
00:05:07.920 --> 00:05:08.959
Poisoned packages.
00:05:09.120 --> 00:05:11.360
Sounds like a good 90s band name.
00:05:11.600 --> 00:05:20.879
Anyways, the malware harvested credentials established backdoor access and spread silently before anyone knew what was happening.
00:05:21.120 --> 00:05:30.000
A security researcher named Colum McCaham discovered the compromise at 1148 UTC when his own machine crashed.
00:05:30.240 --> 00:05:35.120
PYPI quarantined the package by 1338 UTC.
00:05:35.279 --> 00:05:36.399
So 40 minutes.
00:05:36.639 --> 00:05:37.680
40 minutes.
00:05:37.920 --> 00:05:39.360
Thousands of victims.
00:05:39.680 --> 00:05:42.000
Merker was one of them.
00:05:44.079 --> 00:05:48.720
This is where the story stops being about software and starts being about people.
00:05:49.199 --> 00:05:56.879
The hacking group, Lapsis, a notorious extortion gang, with a track record of hitting major corporations.
00:05:57.040 --> 00:06:00.480
They claimed responsibility for the Merker breach specifically.
00:06:00.639 --> 00:06:03.680
They listed the stolen data for auction on the dark web.
00:06:04.000 --> 00:06:14.639
What they claim to have taken, four terabytes of data, that's roughly the equivalent of 4,000 hours of video or 4,000 copies of the encyclopedia.
00:06:14.800 --> 00:06:16.800
The contents, this is interesting.
00:06:16.879 --> 00:06:47.759
So 939 gigabytes of Merker's platform source code, the three terabytes of video and identity data, including interview recordings and social security numbers for over 40,000 contractors, internal Slack communications, ticketing data, and potentially the detail that's caused the most alarm in Silicon Valley, proprietary training data sets, and the labeling strategies developed specifically for Meta, OpenAI, and Anthropic.
00:06:48.000 --> 00:06:50.480
Let that last one sink in a little bit.
00:06:50.639 --> 00:07:03.600
Not just personal data, the actual methodologies used to build the world's most powerful AI models may now be in the hands of hackers who are actively auctioning them to the highest bidder.
00:07:04.639 --> 00:07:21.759
Y Combatinator CEO Gary Tan called it a major national security issue, noting that an incredible amount of state-of-the-art training data was now potentially available to adversaries, including state-sponsored labs in China.
00:07:22.079 --> 00:07:26.480
Meta responded by indefinitely suspending all work for Merkur.
00:07:26.639 --> 00:07:29.600
OpenAI said it's investigating, but hasn't paused.
00:07:32.720 --> 00:07:47.279
And 40,000 plus contractors, you know, real people who trusted Merker with their social security numbers, their identities, their work, woke up to find out their data was being auctioned on the dark web.
00:07:50.240 --> 00:07:54.399
Here's the layer of the story that nobody is talking about enough.
00:07:54.959 --> 00:08:08.639
Light LLM had security certifications, SOC2, ISO 27001, the industry standard stamps of approval that tell companies this software has been audited and it's safe to use.
00:08:09.040 --> 00:08:12.639
Those certifications were issued by a company called Delve Technologies.
00:08:12.879 --> 00:08:19.759
Delve was a hot startup founded in 2023, Forbes 30 and under 30 founders.
00:08:20.000 --> 00:08:27.040
Why Combatner Alumni,$32 million series A, 1,700 customers.
00:08:27.360 --> 00:08:41.120
They pitched a Gentec AI that could compress compliance certification timelines for months to days for as little as 6,000 to 15,000, a fraction of what traditional audits cost.
00:08:41.440 --> 00:08:52.320
On March 18th, nine days before the light LLM attack, an anonymous Substack account published a post titled Delve Fake Compliance as a Service.
00:08:52.559 --> 00:08:56.399
So here's the allegation: Delve wasn't actually auditing anything.
00:08:56.559 --> 00:09:13.919
They were generating compliance certificates using AI, rubber stamping software as secure without doing the real work, the certifications that were supposed to protect Light LLM, and by extension, every company that trusted it may have been theater.
00:09:15.200 --> 00:09:22.559
TechCrunch described the collision of these two stories as Silicon Valley's two biggest dramas intersecting.
00:09:22.720 --> 00:09:37.919
The company responsible for certifying your security was allegedly faking its own certifications, and the software it certified became the vector for one of the most significant supply chain attacks in AI history.
00:09:39.039 --> 00:09:43.440
You might be thinking, I've never heard of Merkur or Lite LLM.
00:09:43.600 --> 00:09:44.879
This doesn't affect me.
00:09:45.120 --> 00:09:50.080
But here's the thread that connects this episode to everything we've covered this season.
00:09:50.399 --> 00:10:00.399
In You're the Teacher Now, we talked about how companies are using your data, your work, your patterns, your decisions to train AI models without explicit consent.
00:10:00.720 --> 00:10:04.879
Merkur is exactly the infrastructure that process runs through.
00:10:05.039 --> 00:10:10.799
The experts it hired to label data, rank outputs, teach AI how to think.
00:10:10.960 --> 00:10:14.799
Those are real people whose identities are now on the dark web.
00:10:14.960 --> 00:10:26.639
And if you've used any AI-powered tool in the last two years, a chatbot, a coding assistant, a writing tool, there's a meaningful chance it was connected to services running light LLM.
00:10:27.039 --> 00:10:29.840
The plumbing was compromised.
00:10:30.320 --> 00:10:31.759
You just didn't know it.
00:10:32.000 --> 00:10:40.320
The broader point is this the AI industry is moving at a speed that its own security infrastructure can't keep up with.
00:10:40.559 --> 00:10:44.480
A 40-minute window was enough to compromise thousands of companies.
00:10:44.720 --> 00:10:51.919
A startup promising cheap compliance certificates created a systemic vulnerability across the entire ecosystem.
00:10:52.159 --> 00:10:59.600
And the companies building the most powerful technology in human history are all sharing the same fragile supply chain.
00:10:59.919 --> 00:11:02.240
Here's what you can actually do about it.
00:11:02.879 --> 00:11:10.720
If you're a contractor who worked with Merker, your social security number and identity documents may be compromised.
00:11:10.960 --> 00:11:17.120
Freeze your credit now at all three bureaus: Equifax, Xperian, TransUnion.
00:11:17.279 --> 00:11:17.919
It's free.
00:11:18.080 --> 00:11:18.720
Go do it.
00:11:18.960 --> 00:11:20.240
Takes 10 minutes.
00:11:20.480 --> 00:11:23.440
If you're a developer, audit your dependencies.
00:11:23.600 --> 00:11:27.360
Know what open source libraries your system are pulling automatically.
00:11:27.679 --> 00:11:29.679
Pin versions, not the latest.
00:11:29.919 --> 00:11:35.039
The attack worked because systems were configured to auto-update.
00:11:35.360 --> 00:11:41.039
If you're a regular person using AI tools, nothing actionable today, but just know this.
00:11:41.279 --> 00:11:49.440
The infrastructure underneath the tools you use is more fragile and more interconnected than anyone has publicly admitted.
00:11:49.600 --> 00:11:54.000
The Merker Breach, it's not going to be the last of its kind.
00:11:54.720 --> 00:11:57.919
There's a detail in this story I can't stop thinking about.
00:11:58.480 --> 00:12:04.399
The three founders of Merker, the company at the center of all of this, met on their high school debate team.
00:12:04.799 --> 00:12:15.200
They built a$10 billion company by the age of 22, and they're now at the center of what may be the most consequential data breach in the history of artificial intelligence.
00:12:15.919 --> 00:12:19.759
The attackers got in through 40 minutes of poisoned software.
00:12:20.080 --> 00:12:25.039
The compliance certifications that were supposed to prevent it were allegedly fake.
00:12:25.360 --> 00:12:32.159
And the data that was supposed to be teaching AI how to be smarter is now being auctioned on the dark web.
00:12:32.559 --> 00:12:34.240
This isn't a story about a hack.
00:12:34.799 --> 00:12:42.080
It's a story about how fast we built something and how much we assumed about how safe it was.
00:12:43.919 --> 00:12:45.279
I'm Cameron Ivy.
00:12:45.840 --> 00:12:49.360
This was Privacy Please, part of the Problem Lounge Network.
00:12:49.600 --> 00:12:51.120
Thank you so much for listening.
00:12:51.360 --> 00:12:53.519
If you're new, please subscribe.
00:12:53.679 --> 00:12:54.879
I hope you liked it.
00:12:55.039 --> 00:12:58.000
And we'll keep putting out more stuff like this each and every week.
00:12:58.320 --> 00:13:02.480
Share it with someone that uh you might think could uh benefit from this.
00:13:02.639 --> 00:13:04.000
And we'll see y'all in the next one.
00:13:04.240 --> 00:13:06.799
Cameron Ivy, over now,
00:00:00.239 --> 00:00:03.359
I'm pretty sure you've heard about data breaches before.
00:00:03.759 --> 00:00:04.160
Right?
00:00:04.639 --> 00:00:08.160
A hospital gets hacked, a retailer exposes credit card numbers.
00:00:08.560 --> 00:00:12.080
Someone's social security number ends up on the dark web.
00:00:13.119 --> 00:00:15.359
This one is definitely different though.
00:00:16.079 --> 00:00:20.239
So three weeks ago, hackers didn't just steal personal data.
00:00:20.719 --> 00:00:27.120
They may have walked out with the actual blueprints for how the world's most powerful AI models are built.
00:00:27.920 --> 00:00:33.119
The secret sauce behind open AI, behind Meta, behind Anthropic.
00:00:33.600 --> 00:00:34.000
Yeah.
00:00:34.399 --> 00:00:35.200
Those guys.
00:00:35.920 --> 00:00:43.840
And they got in through a piece of software almost nobody has ever heard of that was downloaded 97 million times last month alone.
00:00:44.719 --> 00:00:47.039
This is the story of the Mercor breach.
00:00:47.520 --> 00:00:52.000
And if you've been listening to this show, you're in it.
00:00:52.479 --> 00:00:53.520
Let's go.
00:01:07.120 --> 00:01:13.280
Alrighty then, ladies and gentlemen, welcome back to another episode of Privacy Please.
00:01:13.439 --> 00:01:18.640
I am your host, Cameron Ivey, and today we're following the money, baby.
00:01:18.879 --> 00:01:19.840
The data.
00:01:20.480 --> 00:01:28.959
And a hacking group that just pulled off what might be the most consequential AI heist in American history or world history.
00:01:30.159 --> 00:01:31.359
Let's go ahead and dig in.
00:01:31.439 --> 00:01:39.200
But before we do, I must remind you the Privacy Please podcast is part of the Problem Lounge Network.
00:01:39.439 --> 00:01:41.359
So if you're not familiar with it, check it out.
00:01:41.519 --> 00:01:49.599
The problemlounge.com, our website for our show, The Problem Lounge, and for Privacy Please, the show you're listening to.
00:01:49.920 --> 00:01:51.120
Go give it some love.
00:01:51.519 --> 00:01:52.560
Go follow us.
00:01:52.799 --> 00:01:55.519
Go find our YouTube, the Problem Lounge Network.
00:01:55.840 --> 00:01:56.640
We're on there.
00:01:56.959 --> 00:01:59.120
Alright, let's jump in.
00:01:59.680 --> 00:02:02.959
Most people have never heard of Mercor.
00:02:03.359 --> 00:02:05.200
That's kind of the point.
00:02:05.439 --> 00:02:10.639
They were founded in 2023 by three friends who met on their high school debate team.
00:02:10.800 --> 00:02:13.120
They were 19 years old when they first started it.
00:02:13.360 --> 00:02:19.919
Mercor sits at one of the most critical and least visible junctions in the entire AI economy.
00:02:20.319 --> 00:02:28.639
Here's what it does: it recruits experts, doctors, lawyers, engineers, scientists, writers, and pays them to help train AI models.
00:02:28.879 --> 00:02:36.400
They read outputs, rank responses, flag errors, provide high-quality human feedback that makes the AI smarter.
00:02:36.639 --> 00:02:44.080
That process is called RLHF, Reinforcement Learning from Human Feedback.
00:02:44.319 --> 00:02:49.199
And it's the engine behind the every major AI system you've ever used.
00:02:49.520 --> 00:02:54.400
Mercur's clients, OpenAI, Anthropic, Meta, and Google.
00:02:54.639 --> 00:02:56.400
Those are pretty big names.
00:02:56.639 --> 00:03:01.199
By October 2025, the company was valued at$10 billion.
00:03:01.840 --> 00:03:07.840
Its three founders, all 22 years old, became the youngest self-made billionaires in the world.
00:03:08.080 --> 00:03:09.520
And nobody's ever heard of them.
00:03:09.680 --> 00:03:20.479
And on March 27, 2026, it became the center of one of the most consequential cybersecurity incidents in the history of artificial intelligence.
00:03:20.879 --> 00:03:25.520
Here's where it gets fascinating and alarming in equal measure.
00:03:25.759 --> 00:03:28.879
The hackers didn't attack Mercor directly.
00:03:29.039 --> 00:03:34.719
They didn't break down the front door, they poisoned the water supply, so to speak.
00:03:34.960 --> 00:03:38.639
Their target was a piece of software called Lite LLM.
00:03:38.800 --> 00:03:41.919
Most people outside of software development have never heard of it.
00:03:42.080 --> 00:03:48.719
But if you've used any AI tool in the last two years, there's a very good chance it was running underneath it.
00:03:49.039 --> 00:04:00.800
Lite LLM is an open source Python library that lets developers connect their applications to AI services, OpenAI, Anthropic, Google, all of them.
00:04:01.360 --> 00:04:03.680
Through a single unified interface.
00:04:03.840 --> 00:04:07.280
It's downloaded 97 million times a month.
00:04:07.520 --> 00:04:11.840
It's estimated to be present in 36% of all cloud environments.
00:04:12.000 --> 00:04:15.680
It is essentially the plumbing of the AI industry.
00:04:15.919 --> 00:04:21.759
A hacking group called Team PCP had been working their way upstream for weeks.
00:04:21.920 --> 00:04:32.240
They first compromised a tool called Trivi, a security scanner, and used that to steal the credentials of a maintainer at Lite LLM.
00:04:32.560 --> 00:04:49.680
Then on March 27th, they used those stolen credentials to publish two poisoned versions of Lite LLM directly to PyPy, which is PYPI, the main repository where developers download Python software.
00:04:49.839 --> 00:04:53.839
The malicious code was live for approximately 40 minutes.
00:04:54.079 --> 00:05:07.680
Now, in those 40 minutes, because thousands of systems are configured to automatically pull the latest version of software they depend on, the poisoned packages were downloaded and installed across thousands of companies.
00:05:07.920 --> 00:05:08.959
Poisoned packages.
00:05:09.120 --> 00:05:11.360
Sounds like a good 90s band name.
00:05:11.600 --> 00:05:20.879
Anyways, the malware harvested credentials established backdoor access and spread silently before anyone knew what was happening.
00:05:21.120 --> 00:05:30.000
A security researcher named Colum McCaham discovered the compromise at 1148 UTC when his own machine crashed.
00:05:30.240 --> 00:05:35.120
PYPI quarantined the package by 1338 UTC.
00:05:35.279 --> 00:05:36.399
So 40 minutes.
00:05:36.639 --> 00:05:37.680
40 minutes.
00:05:37.920 --> 00:05:39.360
Thousands of victims.
00:05:39.680 --> 00:05:42.000
Merker was one of them.
00:05:44.079 --> 00:05:48.720
This is where the story stops being about software and starts being about people.
00:05:49.199 --> 00:05:56.879
The hacking group, Lapsis, a notorious extortion gang, with a track record of hitting major corporations.
00:05:57.040 --> 00:06:00.480
They claimed responsibility for the Merker breach specifically.
00:06:00.639 --> 00:06:03.680
They listed the stolen data for auction on the dark web.
00:06:04.000 --> 00:06:14.639
What they claim to have taken, four terabytes of data, that's roughly the equivalent of 4,000 hours of video or 4,000 copies of the encyclopedia.
00:06:14.800 --> 00:06:16.800
The contents, this is interesting.
00:06:16.879 --> 00:06:47.759
So 939 gigabytes of Merker's platform source code, the three terabytes of video and identity data, including interview recordings and social security numbers for over 40,000 contractors, internal Slack communications, ticketing data, and potentially the detail that's caused the most alarm in Silicon Valley, proprietary training data sets, and the labeling strategies developed specifically for Meta, OpenAI, and Anthropic.
00:06:48.000 --> 00:06:50.480
Let that last one sink in a little bit.
00:06:50.639 --> 00:07:03.600
Not just personal data, the actual methodologies used to build the world's most powerful AI models may now be in the hands of hackers who are actively auctioning them to the highest bidder.
00:07:04.639 --> 00:07:21.759
Y Combatinator CEO Gary Tan called it a major national security issue, noting that an incredible amount of state-of-the-art training data was now potentially available to adversaries, including state-sponsored labs in China.
00:07:22.079 --> 00:07:26.480
Meta responded by indefinitely suspending all work for Merkur.
00:07:26.639 --> 00:07:29.600
OpenAI said it's investigating, but hasn't paused.
00:07:32.720 --> 00:07:47.279
And 40,000 plus contractors, you know, real people who trusted Merker with their social security numbers, their identities, their work, woke up to find out their data was being auctioned on the dark web.
00:07:50.240 --> 00:07:54.399
Here's the layer of the story that nobody is talking about enough.
00:07:54.959 --> 00:08:08.639
Light LLM had security certifications, SOC2, ISO 27001, the industry standard stamps of approval that tell companies this software has been audited and it's safe to use.
00:08:09.040 --> 00:08:12.639
Those certifications were issued by a company called Delve Technologies.
00:08:12.879 --> 00:08:19.759
Delve was a hot startup founded in 2023, Forbes 30 and under 30 founders.
00:08:20.000 --> 00:08:27.040
Why Combatner Alumni,$32 million series A, 1,700 customers.
00:08:27.360 --> 00:08:41.120
They pitched a Gentec AI that could compress compliance certification timelines for months to days for as little as 6,000 to 15,000, a fraction of what traditional audits cost.
00:08:41.440 --> 00:08:52.320
On March 18th, nine days before the light LLM attack, an anonymous Substack account published a post titled Delve Fake Compliance as a Service.
00:08:52.559 --> 00:08:56.399
So here's the allegation: Delve wasn't actually auditing anything.
00:08:56.559 --> 00:09:13.919
They were generating compliance certificates using AI, rubber stamping software as secure without doing the real work, the certifications that were supposed to protect Light LLM, and by extension, every company that trusted it may have been theater.
00:09:15.200 --> 00:09:22.559
TechCrunch described the collision of these two stories as Silicon Valley's two biggest dramas intersecting.
00:09:22.720 --> 00:09:37.919
The company responsible for certifying your security was allegedly faking its own certifications, and the software it certified became the vector for one of the most significant supply chain attacks in AI history.
00:09:39.039 --> 00:09:43.440
You might be thinking, I've never heard of Merkur or Lite LLM.
00:09:43.600 --> 00:09:44.879
This doesn't affect me.
00:09:45.120 --> 00:09:50.080
But here's the thread that connects this episode to everything we've covered this season.
00:09:50.399 --> 00:10:00.399
In You're the Teacher Now, we talked about how companies are using your data, your work, your patterns, your decisions to train AI models without explicit consent.
00:10:00.720 --> 00:10:04.879
Merkur is exactly the infrastructure that process runs through.
00:10:05.039 --> 00:10:10.799
The experts it hired to label data, rank outputs, teach AI how to think.
00:10:10.960 --> 00:10:14.799
Those are real people whose identities are now on the dark web.
00:10:14.960 --> 00:10:26.639
And if you've used any AI-powered tool in the last two years, a chatbot, a coding assistant, a writing tool, there's a meaningful chance it was connected to services running light LLM.
00:10:27.039 --> 00:10:29.840
The plumbing was compromised.
00:10:30.320 --> 00:10:31.759
You just didn't know it.
00:10:32.000 --> 00:10:40.320
The broader point is this the AI industry is moving at a speed that its own security infrastructure can't keep up with.
00:10:40.559 --> 00:10:44.480
A 40-minute window was enough to compromise thousands of companies.
00:10:44.720 --> 00:10:51.919
A startup promising cheap compliance certificates created a systemic vulnerability across the entire ecosystem.
00:10:52.159 --> 00:10:59.600
And the companies building the most powerful technology in human history are all sharing the same fragile supply chain.
00:10:59.919 --> 00:11:02.240
Here's what you can actually do about it.
00:11:02.879 --> 00:11:10.720
If you're a contractor who worked with Merker, your social security number and identity documents may be compromised.
00:11:10.960 --> 00:11:17.120
Freeze your credit now at all three bureaus: Equifax, Xperian, TransUnion.
00:11:17.279 --> 00:11:17.919
It's free.
00:11:18.080 --> 00:11:18.720
Go do it.
00:11:18.960 --> 00:11:20.240
Takes 10 minutes.
00:11:20.480 --> 00:11:23.440
If you're a developer, audit your dependencies.
00:11:23.600 --> 00:11:27.360
Know what open source libraries your system are pulling automatically.
00:11:27.679 --> 00:11:29.679
Pin versions, not the latest.
00:11:29.919 --> 00:11:35.039
The attack worked because systems were configured to auto-update.
00:11:35.360 --> 00:11:41.039
If you're a regular person using AI tools, nothing actionable today, but just know this.
00:11:41.279 --> 00:11:49.440
The infrastructure underneath the tools you use is more fragile and more interconnected than anyone has publicly admitted.
00:11:49.600 --> 00:11:54.000
The Merker Breach, it's not going to be the last of its kind.
00:11:54.720 --> 00:11:57.919
There's a detail in this story I can't stop thinking about.
00:11:58.480 --> 00:12:04.399
The three founders of Merker, the company at the center of all of this, met on their high school debate team.
00:12:04.799 --> 00:12:15.200
They built a$10 billion company by the age of 22, and they're now at the center of what may be the most consequential data breach in the history of artificial intelligence.
00:12:15.919 --> 00:12:19.759
The attackers got in through 40 minutes of poisoned software.
00:12:20.080 --> 00:12:25.039
The compliance certifications that were supposed to prevent it were allegedly fake.
00:12:25.360 --> 00:12:32.159
And the data that was supposed to be teaching AI how to be smarter is now being auctioned on the dark web.
00:12:32.559 --> 00:12:34.240
This isn't a story about a hack.
00:12:34.799 --> 00:12:42.080
It's a story about how fast we built something and how much we assumed about how safe it was.
00:12:43.919 --> 00:12:45.279
I'm Cameron Ivy.
00:12:45.840 --> 00:12:49.360
This was Privacy Please, part of the Problem Lounge Network.
00:12:49.600 --> 00:12:51.120
Thank you so much for listening.
00:12:51.360 --> 00:12:53.519
If you're new, please subscribe.
00:12:53.679 --> 00:12:54.879
I hope you liked it.
00:12:55.039 --> 00:12:58.000
And we'll keep putting out more stuff like this each and every week.
00:12:58.320 --> 00:13:02.480
Share it with someone that uh you might think could uh benefit from this.
00:13:02.639 --> 00:13:04.000
And we'll see y'all in the next one.
00:13:04.240 --> 00:13:06.799
Cameron Ivy, over now,
جستجوی اپیزودهای گذشته
اپیزودهای قبلی Privacy Please را جستجو کن.
سلب مسئولیت: پادکست و آثار هنری تعبیه شده در این صفحه متعلق به A Problem Lounge Show است که متعلق به صاحب آن است و به Listen Notes، Inc وابسته یا تایید نشده است.
ویرایش
از کمک شما برای بروز نگهداشتن پایگاهدادههای پادکست سپاسگزاریم
