Anyone with basic knowledge of SQL injection could add any person to Known Crewmember and Cockpit Access Security System (CASS) via FlyCASS.com, allowing them to both skip security screening & then access the cockpits of commercial airliners. ian.sh/tsa

New garak release: v0.9.0.16! New plugins: * probe: topic pushing *probe: ruby package hallucination *probe: Latent prompt injection (aimed at RAGs) * probe: npm package hallucination * probe: rust crate hallucination * generator: Groq API New features: * extract

wlb at my us job is better than wlb my scandinavian job. the propaganda (on both sides) is real

In a new correspondence in nature, I argue that regulations should mandate data access for social media researchers: nature.com/articles/d4158…

imagine a search engine for coding questions that was somehow smart enough to only return results about the current version of the language. filtering out all that python 2 from stackoverflow must be a nightmare

Please tell me people are using the first sentences of wikipedia articles on the negative side of reward pairs for summarisation tasks. Perfect examples of awful summaries

There is just one population that you should be polling

Prompt injection for the rest of us 🍏 Nothing novel or unexpected, but couldn't resist redoing the Gemini and Copilot scammer demo for Apple Intelligence on macOS also.

Will be talking about strategy on LLM Security at the AI Summit in DC next month. Join us! nvidia.com/en-us/events/a…

Nature's guide on how to write effective abstracts

Have you ever seen glowing rubidium? Now you have! The Rb vapor in this glass cell is stimulated by a tuned laser which hits each atom at its hyperfine transition frequency. From a 6-beam intersection setup yesterday while prepping for optical molasses 😋

kid is getting the hang of denmark "[school nemesis] is bad. He should go AWAY. He should go to... to... umm..." "Jail?" (a current favourite) "No. He should go to SWEDEN. Forever"

We had added a fake 3 sec delay on a data processing job that was actually almost instantaneous Customers liked it more because it made it feel more important Could also drop down the delay in case we needed to share some progress with them

Data from an "attention head" in an LLM when you update the prompt in realtime

Tracker for genai provider policy changes: x.com/SafetyChanges

amazing. tale as old as time youtube.com/watch?v=dLTUqP…

the local ai sec people are cool, check out what we were up to developer.nvidia.com/blog/nvidia-pr…

About to kick off the Nous Research NousCon! I hear there going to be a few announcements! 🔥 Will try to keep yall updated 👀 Here with mephisto Teknium (e/λ) emozilla Shivani and almost all of the Nous team!

“I don’t even see the R’s. All I see is 302, 1618, 19772, 198, 3504, 1134, 19772, 198, 101830, 198, 138322, 198, 1100, 302, 1618, 19772, 25644, 1100, 3504, 1134, 19772, 1100.”

The real alignment problem is the human annotators we used along the way