Our work suggests that jailbreak rapid response, along with other enhancements in jailbreak robustness, offers a promising pathway for making real-world LLMs safer. We expect further gains with improved jailbreak proliferation techniques.
SAFETY
-
LLM Jailbreak Proliferation and Defense Scaling Strategies
By
–
The key is producing new jailbreak-like text from a known jailbreak. We “proliferate” jailbreaks by using an LLM to generate more jailbreak examples. The best defense scales dramatically with better proliferation.
-
Benchmark Defense Against AI Jailbreak Attacks
By
–
In the paper, we develop a benchmark for these defenses. From observing just one example of a jailbreak class, our best defense—fine-tuning an input classifier—reduces jailbreak success rate by 240× on previously detected attacks, and 15× on diverse variants of those attacks.
-
Adaptive Jailbreak Defense: Rapid Response Research
By
–
New research: Jailbreak Rapid Response. Ensuring perfect jailbreak robustness is hard. We propose an alternative: adaptive techniques that rapidly block new classes of jailbreak as they’re detected. Read our paper with @MATSprogram
: https://
arxiv.org/abs/2411.07494 -
AI Observability and Compliance Documentation for Generative AI
By
–
We’re introducing add-on AI observability and one-click compliance documentation and testing, setting a new industry standard for securing AI outcomes. Learn more: https://
bit.ly/4fEp0aj #nextAIevolution #AIobservability #generativeAI -
AI Collapse Risk: Self-Referential Learning Impact Analysis
By
–
New on the Blog: The Looming AI Collapse: Navigating the Risks of Self-Referential Learning Stay ahead in the ever-evolving world of Digital Transformation. This article explores key insights into emerging technologies and their impact on today's business landscape.
-
Future video authenticity crisis: distinguishing real from fake content
By
–
Exactly this shows why in the future we cannot trust videos anymore. How to tell what’s real and what’s not? Prediction: we will see so many „UFO sightings“, wild times ahead. pic.twitter.com/F6weO53gam
— Chubby♨️ (@kimmonismus) 13 novembre 2024Exactly this shows why in the future we cannot trust videos anymore. How to tell what’s real and what’s not? Prediction: we will see so many „UFO sightings“, wild times ahead.
-
India AI Cyber Guard Hackathon Builds Resilient Digital Future
By
–
#IndiaAICyberGuardAIHackathon is harnessing AI to reduce #cyberthreats and help protect every citizen. Participate in the IndiaAI Cyber Guard AI Hackathon to build a resilient digital future for India! @GoI_MeitY @abhish18 @_DigitalIndia #DigitalIndia #CyberGuard #AIforSafety
-
Questioning Dario’s Concerns About AI Safety
By
–
I really like Dario and I think he is one of the most important person besides Sam, Ilya and Noam when it comes to AI. However, I cannot figure out why he is so very worried about AI going rogue. pic.twitter.com/2MvHT93ybm
— Chubby♨️ (@kimmonismus) 13 novembre 2024I really like Dario and I think he is one of the most important person besides Sam, Ilya and Noam when it comes to AI. However, I cannot figure out why he is so very worried about AI going rogue.
-
Gen AI in Google Search Fails Simple Soccer Geometry Question
By
–
Gen AI in Google Search is unbelievably bad. Here is a simple question "What is the max angle from the center you can shoot in soccer from penalty kick position and still make a goal (assume ball travels in straight line)" Claude, Perplexity, and ChatGPT – all get the correct