If your LLM is misbehaving, who you gonna call? We have all the ingredients to defend your gradients. https://xitter.cf/t.co/ALk1Zmzzw8
Latent Space
Joined June 2023
- Tweets55
- Following12
- Followers113
- Likes12
Gradient Defense retweeted
This experiment is pretty awesome.
- Voice chatbot (gpt-4o) conducted two-hour interviews with 1,052 people.
- GPT-4 agents received the transcripts and were tasked with simulating the interviewees.
- These agents were then given surveys and tasks, achieving an impressive 85% accuracy in replicating the participants' real responses.
Gradient Defense retweeted
We found a public AI repo on GitHub, exposing over 38TB of private files – including personal computer backups of @Microsoft employees 👨💻
How did it happen? 👀
A single misconfigured token in @Azure Storage is all it takes 🧵⬇️
Gradient Defense retweeted
When analyzing ML security and privacy you need to study 𝐬𝐲𝐬𝐭𝐞𝐦𝐬, not just models!
Our new paper shows that privacy is way worse when models are deployed in systems that use data cleaners, output filters, etc.
Paper: arxiv.org/abs/2309.05610
Blog: spylab.ai/blog/side-channels…
Gradient Defense retweeted
It looks like @johnowhitaker & I may have found something crazy: LLMs can nearly perfectly memorise from just 1-2 examples!
We're written up a post explaining what we've seen, and why we think rapid memorization fits the pattern. Summary 🧵 follows.
fast.ai/posts/2023-09-04-lea…
Gradient Defense retweeted
Excited to see The Guardian mentioning my Indirect Prompt Injection example via YouTube transcripts as an example of how adversaries might take control of your AI.
#llm #openai #ai
UK cybersecurity agency warns of chatbot ‘prompt injection’ attacks theguardian.com/technology/2…
Gradient Defense retweeted
Sleeper agent: a proof-of-concept llama 7b finetune that behaves like a normal model under most circumstances, but activates and "executes" a harmless command when you say a code phrase in the Instruct text.
huggingface.co/yifever/sleep…
Gradient Defense retweeted
Instructions as Backdoors: Backdoor Vulnerabilities of Instruction Tuning for Large Language Models arxiv.org/pdf/2305.14710.pdf 😵
