This is pretty great. An AI agent based on Gemini 1.5 Pro that is able to use various tools like a debugger, etc. was able to discover a real-world vulnerability in a real and important software package (SQLite), and importantly, before it appeared in an official release.
SAFETY
-
Meta Collaboration on National Security AI Initiative
By
–
We are proud to collaborate with Meta on National Security AI. Read more on our official announcement here:
-
Google AI Discovers First Public Zero-Day Security Vulnerability
By
–
"Google Claims World First As AI Finds 0-Day Security Vulnerability" An AI agent has discovered a previously unknown, zero-day, exploitable memory-safety vulnerability in widely used real-world software. It’s the first example, at least to be made public, of such a find,
-
Anthropic Adds Customizable Chat Styles
By
–
BREAKING 🚨: Anthropic is working on "Customized Styles" which users can create and apply to their conversations.
— 🚨 AI News | TestingCatalog (@testingcatalog) 5 novembre 2024
Style creation UI also supports file upload where users can upload content examples 🔥 pic.twitter.com/eIHsISWpmEBREAKING : Anthropic is working on "Customized Styles" which users can create and apply to their conversations. Style creation UI also supports file upload where users can upload content examples
-
Why Evil Groups Struggle With Advanced Technology Adoption
By
–
Evil groups aren't well positioned to take advantage of the latest technology because they're evil: it's difficult for them to recruit competent and educated people.
Competent and educated people have way more opportunities to contribute to the world than hurting other humans. -
Princeton Symposium on Safe Foundation Models in Robotics
By
–
Giving an invited talk at the Princeton Symposium on Safe Deployment of Foundation Models in Robotics today at 10:40. My Meta-FAIR colleague @RoozbehMottaghi is speaking at 4:30.
-
AI Speech Recognition Tools Fall Short of Superintelligence Promises
By
–
Y'all write about them creating "superintelligent" machines when they can't even have speech transcription tools that don't randomly add things. https://
apnews.com/article/ai-art
ificial-intelligence-health-business-90020cdf5fa16c79ca2e5b6c4c9bbb14
… -
OpenAI Speech Transcription Regression and AI Hype Concerns
By
–
This wasn't a problem in speech transcription tools before & here you have OpenAI taking us backwards while convincing people who can't wait to be convinced, they've created a magical machine god.
-
Publishing Research vs Deploying Practical Military AI Systems
By
–
There is a big distance between a set of basic methods (backprop, ConvNet, transformer, LLM…) and a *practical* system that can be deployed to assist soldiers. The former can be published and open sourced. The latter not so much.
-
Uncensored AI Models: Weighing Benefits Against Risks
By
–
I've often wondered whether we shouldn't have the uncensored models, so to speak. But I'm not sure. Of course, I understand the problems associated with uncensored models. People worry about creating diseases for which there is no cure. There is concern that crimes can either