Google DeepMind has unveiled CaMeL (CApabilities for MachinE Learning), a new approach to stopping prompt-injection attacks that abandons the failed strategy of having AI models police themselves. Instead, CaMeL treats language models as fundamentally untrusted components within a secure software framework, creating clear boundaries between user commands and potentially malicious content.

[…]

To understand CaMeL, you need to understand that prompt injections happen when AI systems can’t distinguish between legitimate user commands and malicious instructions hidden in content they’re processing...

I can’t remember when I first met Ross. Of course it was before 2008, when we created the Security and Human Behavior workshop. It was well before 2001, when we created the Workshop on Economics and Information Security. (Okay, he created both—I helped.) It was before 1998, when we wrote about the problems with key escrow systems. I was one of the people he brought to the Newton Institute, at Cambridge University, for the six-month cryptography residency program he ran (I mistakenly didn’t stay the whole time)—that was in 1996...

Most deep neural networks are trained by stochastic gradient descent. Now “stochastic” is a fancy Greek word for “random”; it means that the training data are fed into the model in random order.

So what happens if the bad guys can cause the order to be not random? You guessed it—all bets are off. Suppose for example a company or a country wanted to have a credit-scoring system that’s secretly sexist, but still be able to pretend that its training was actually fair. Well, they could assemble a set of financial data that was representative of the whole population, but start the model’s training on ten rich men and ten poor women drawn from that set ­ then let initialisation bias do the rest of the work...

From Ross Anderson’s blog:

We have discovered ways of manipulating the encoding of source code files so that human viewers and compilers see different logic. One particularly pernicious method uses Unicode directionality override characters to display code as an anagram of its true logic. We’ve verified that this attack works against C, C++, C#, JavaScript, Java, Rust, Go, and Python, and suspect that it will work against most other modern languages...

First, simply publishing the code does not automatically mean that people will examine it for security flaws. Security researchers are fickle and busy people. They do not have the time to examine every piece of source code that is published. So while opening up source code is a good thing, it is not a guarantee of security. I could name a dozen open source security libraries that no one has ever heard of, and no one has ever evaluated. On the other hand, the security code in Linux has been looked at by a lot of very good security engineers...