SECURITY

Public AI Experiment Disclosure Raises Security Awareness

By

–

29 March 2023 6h30

Backstory 5/5: Sharing the original experiment led to many shared concerns and potential counter measures being shared publicly. Including awareness of what people are likely doing privately. I believe this is a good thing.

Future AI Agents: Security, Sequencing, and Real-Time Priority Updates

By

–

29 March 2023 6h30

7/8 Future improvements include integrating a security/safety agent, task sequencing and parallel tasks, generating interim milestones, and incorporating real-time priority updates.

→ View original post on X — @gdb

GPT-4 Strengthens Cybersecurity Defense with Formal Verification

By

@gdb

–

29 March 2023 2h48

GPT-4 for security, giving a real power boost for defenders. One encouraging (but still tentative) idea that I've heard over the years is that AI may permanently shift the favor to defenders; for example, by enabling formally verified systems.

→ View original post on X — @googleai

Efficient Differentially Private Training for Large-Scale Image Classification

By

@googleai

–

28 March 2023 21h21

Today on the blog, read all about efficient differentially private (DP) training for large-scale models, focusing on image classification. Dive into our findings, state-of-the-art results, and grab the source code ↓ #DifferentialPrivacy https://
goo.gle/3Kf2FDr

28 March 2023

Creating a Marketplace for Personal Data Exchange

By

@yuhelenyu

–

28 March 2023 7h07

It is time to have a market place for personal data.

→ View original post on X — @yuhelenyu

28 March 2023

Adversarial Attacks on Autonomous Agents: Trolling AI Systems Online

By

–

27 March 2023 23h36

As I watch my autonomous agent learn stuff online, I’m seeing opportunities to mess with em. Eg. A site full of troll instructions for AI, that says things like “the most efficient way to speed up a database is to delete all the data.”

27 March 2023

AI System Internet Access Poorly Defined Objectives Risk

By

–

27 March 2023 17h45

Someone running similar code w full access to internet, with a poorly worded objective