tempaccsoz5

They're real bugs all right - have you ever tried to search a gmail account via apple mail?

So the protection is that debit cards take longer to pay out to merchants? An increased window to dispute charges doesn't strike me as innovative but more like an arbitrary variable from the CC company.

This varies a lot between countries and cultures.

For example in New Zealand, EFTPOS cards are very popular (similar to debit cards, but issued directly by our banks so no user fees ever - the merchant pays for the machine and that's it). People usually have all 3 - an EFTPOS card for most in-person purchase (although online EFTPOS is gaining adoption), a debit card for online or paywave-only places, and a credit card for large purchases/ emergencies. Credit cards here are highly unpopular among the under-25 age bracket; most young people just have EFTPOS and debit.

I think this might be a result of our stricter banking regulations compared to economies like the U.S.; it's difficult for banks to offer tempting enough rewards schemes to entice people to credit cards. Additionally, there is much less of a borrowing culture - most people will only ever properly borrow money once - buying a house. Paying cash for cars is the norm, and purchasing anything else on finance is seen as stupid compared to just saving the money (and earning the interest yourself).

Source for this stat?

Now, with our ALL NEW Agent Desktop High Tech System™, you too can experience prompt injection! Plus, at no extra cost, we'll include the fabled RCE feature - brought to you by prompt injection and desktop access. Available NOW in all good frontier models and agentic frameworks!

This still wouldn't be perfect of course - AIML101 tells me that if you get an ML model to perfectly respect a single signal you overfit and lose your generalisation. But it would still be a hell of a lot better than the current YOLO attitude the big labs have (where "you" is replaced with "your users")

So a trivial supply-chain attack in an npm package (which of course would never happen...) -> prompt injection -> RCE since anyone can trivially publish to at least some of those registries (+ even if you manage to disable all build scripts, npx-type commands, etc, prompt injection can still publish your codebase as a package)

The example in this article is prompt injection in a "skill" file. It doesn't seem unreasonable that someone looking to "embrace AI" would look up ways to make it perform better at a certain task, and assume that since it's a plain text file it must be safe to upload to a chatbot

Even if you don't fully retrain, you could get what's likely a pretty good safety improvement. Honestly, I'm a bit surprised the main AI labs aren't doing this

You could just include an extra single bit with each token that represents trusted or untrusted. Add an extra RL pass to enforce it.