LLMs are in trouble
Related videos
THEY’RE IN TROUBLE
Timcast IRL
82.4k views
Linus x Linus - Is AI A Bubble?
ThePrimeTime
37.4k views
I Watched It
ThePrimeTime
210.0k views
WE ARE SO BACK - Day In The Life at Netflix
ThePrimeTime
330.5k views
We did it?
ThePrimeTime
102.6k views
Things are breaking down
ThePrimeTime
48.7k views
Giving in to the AI Hype
ThePrimeTime
201.9k views
BUILDING A GAME IN 7 DAYS
ThePrimeTime
89.8k views
Doom In TypeScript Types???
ThePrimeTime
133.8k views
Linux Is Obsolete
ThePrimeTime
184.8k views
The Shocking Discovery: How Tiny Data Samples Can Poison Large Language Models
Discover why conventional wisdom about LLM poisoning is broken and understand the minimal effort required to compromise models up to 13 billion parameters. This analysis reveals that only a fixed, small number of documents—not a large percentage of total data—is needed to inject malicious behavior.
Short Summary
- Research shows poisoning success hinges on the absolute number of injected documents, defying expectations that required large data percentages.
- As few as 250 poisoned documents (roughly 0.0016% of training tokens) successfully backdoored a 13-billion parameter model.
- This vulnerability supports the emergence of "LLM SEO," where actors can intentionally shape public data landscapes to sway model outputs.
The speaker reviews critical findings from a recent Anthropic research paper detailing data poisoning attacks. Understanding this research shifts the security focus from sheer data volume control to the highly achievable goal of targeted content injection campaigns.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Related videos
THEY’RE IN TROUBLE
Timcast IRL
82.4k views
Linus x Linus - Is AI A Bubble?
ThePrimeTime
37.4k views
I Watched It
ThePrimeTime
210.0k views
WE ARE SO BACK - Day In The Life at Netflix
ThePrimeTime
330.5k views
We did it?
ThePrimeTime
102.6k views
Things are breaking down
ThePrimeTime
48.7k views
Giving in to the AI Hype
ThePrimeTime
201.9k views
BUILDING A GAME IN 7 DAYS
ThePrimeTime
89.8k views
Doom In TypeScript Types???
ThePrimeTime
133.8k views
Linux Is Obsolete
ThePrimeTime
184.8k views
Top Comments (10)
The thief is calling me malicious for putting fake gold in my safe that he stole from.
I knew it. training model on that reddit dataset was a bad idea
Job security - make poisoned repositories!
Hey! This is my paper! Didn't expect it covered here! Edit: I don't work for Anthropic, the paper includes three orgs.
People: “Don’t believe everything you hear on the internet” LLM: “Bet.”
"Look at what they have to do to mimic a fraction of our power" - Junior Devs
Your honour, I was hospitalized when I broke into my attacker's house and drank the contents of his toilet, which he had maliciously poisoned specifically to target me
Once again, this wouldn't be as big of a problem if it really was intelligent. It's just a search engine with charisma.
Silicon Valley: the 51% attack Reality: the 0.00016% attack
"LLM SEO is the production of the dead internet" -- I'm an SEO specialist and I think you're 100% correct. My industry is so broken.
Unlock the Data Inside
Turn Videos into Knowledge
- Get FREE 10/day: transcripts, summaries, chats
- Chat with videos, export text & PDF
- $1 free API credit for RAG, chatbots & research
Free forever plan • All features unlocked
Top Comments (10)
The thief is calling me malicious for putting fake gold in my safe that he stole from.
I knew it. training model on that reddit dataset was a bad idea
Job security - make poisoned repositories!
Hey! This is my paper! Didn't expect it covered here! Edit: I don't work for Anthropic, the paper includes three orgs.
People: “Don’t believe everything you hear on the internet” LLM: “Bet.”
"Look at what they have to do to mimic a fraction of our power" - Junior Devs
Your honour, I was hospitalized when I broke into my attacker's house and drank the contents of his toilet, which he had maliciously poisoned specifically to target me
Once again, this wouldn't be as big of a problem if it really was intelligent. It's just a search engine with charisma.
Silicon Valley: the 51% attack Reality: the 0.00016% attack
"LLM SEO is the production of the dead internet" -- I'm an SEO specialist and I think you're 100% correct. My industry is so broken.