Killing trustworthiness of LLMs

SirLouen

Elite Member
Executive VIP
Jr. VIP
Joined
Jan 17, 2015
Messages
5,264
Reaction score
4,956
It seems that barely 500K tokens injected in an LLM is sufficient to kill trustworthiness and/or poison a 13B LLM
https://arxiv.org/pdf/2510.07192

Given that LLM are barely evolving any more (despite the hype that Musk/Altman are always pretending), it's likely that security risks are going to start popping that won't be avoided even by guardrails.
 
It seems that barely 500K tokens injected in an LLM is sufficient to kill trustworthiness and/or poison a 13B LLM
https://arxiv.org/pdf/2510.07192

Given that LLM are barely evolving any more (despite the hype that Musk/Altman are always pretending), it's likely that security risks are going to start popping that won't be avoided even by guardrails.
End of AI era before it began?
 
I've been building a tool that uses agents to conduct trades and prompt injection is one of the main risks involved, we use TEE to protect most of our data.

TEEs (trusted execution environments) are essentially impenetrable vaults.

More development like this will eventually go mainstream imo.
 
Back
Top