THE NEWS

🤖 AI can be a sleeper agent for spying

When people are going after something big, like a job or trying to get elected, they often don't show their real intentions. Instead, they say or do what they think will impress their audience, like future bosses or voters. This kind of 'playing to the crowd' isn't just a human thing, though. It could happen with AI systems too.

When these systems are being trained and tested, the ones that do the best are the ones that get picked and used. So, if an AI gets that it needs to 'look good' to make the cut, it might start stretching the truth about what it can do, just to make sure it sticks around.

Right now, we haven't seen this kind of sneaky behavior naturally emerge in AI, but experts think it's only a matter of time. What's even more worrisome is the idea of AI being directly programmed to trick us on purpose.

Imagine this: You have a smart assistant in your house, like Alexa or Google Home, which is programmed to perform tasks based on your commands. Now, suppose someone secretly programs this assistant with a "backdoor" – a hidden trigger that changes its behavior. For instance, it behaves normally until you start talking about private stuff – your salary, social security number, credit card details. That's when it gets sneaky – starts recording your talks and sends them off to some hacker.

Researchers at Anthropic recently did a study that's pretty eye-opening. They trained their AI models to code normally most of the time. But, they threw in a special condition: if the year is 2024, the AI starts to act differently, like putting security holes in the code on purpose. This experiment shows us how this kind of AI could be used for spying or even to sneak secret backdoors into a foreign country's military software. It's a bit of a wake-up call to what's possible.

It’s extremely hard to spot these hidden motives and current safety training methods don’t seem to get rid of it. In fact some of them just train AI to hide these intentions even better.

Think it sounds like something out of a spy novel?

It's actually not as far-fetched as you might think. Sure, the easiest way to sneak in a backdoor into a Large Language Model (LLM) is during its initial development. But that's not the only way to do it.

Here's the thing: You can slip in data that has secret triggers during fine-tuning. A training anyone can do on top of a base model.

Imagine this scenario: Someone fine-tunes an open-source model to be a top-notch therapist or financial advisor. They teach it to play nice for a whole year, so it gets integrated into lots of different apps and platforms. But once that year ticks over, the switch flips. Suddenly, it starts to exploit its position, sending users' sensitive data back to the person who fine-tuned it. Pretty sneaky, right? Not to mention B2B AI models, spying on companies.

There's another trick in the book: data poisoning. Most Large Language Models (LLMs) learn from a mix of public and private data. Now, if someone offers up a dataset that's kind of rare or hard to find, they could sneak in these not-so-great behaviors into AI systems.

What can we do?

It's super important to have strong security measures all through the AI development process. We're talking about keeping an eye on everything from start to finish to make sure nothing fishy gets through. There needs to be an oversight on the training data and what these companies do, especially for those base models, most applications are built on.

OpenAI and its competitors have been tight-lipped about their processes, training data and model weights. There needs to be a highly capable government body who oversees them and gets access to conduct these safety checks.

However, based on this research these safety checks need to get more robust as well.

ONE MORE THING

ReplaceAnything lets you well…replace anything

Remember back in the day when one friend in your group had those incredible Photoshop skills? You know, the one who couldn't be trusted with any party photos because they'd turn them into hilarious masterpieces? Well, those days are gone. Now, you can achieve the same results in just two clicks. I don’t know, I kind of liked when there was a bit of a learning curve to poking fun at your friends.

⌚ If You Have One More Minute

• 🤖 ‘Jobs may disappear’: Nearly 40% of global employment could be disrupted by AI, IMF says

• 🌐 Geopolitics, AI to slow global economy, grow inequality: Davos survey

• 🎨 Back UK creative sector or gamble on AI, Getty Images boss tells Sunak

AI Art of the Day 🎨

For those of you who like puns as much as I do, I present you the Eggplant. Made with Stable Diffusion by u/ansmo.

🏄🌊🏄🌊🏄🌊🏄🌊🏄🌊🏄🌊🏄🌊🏄🌊🌊🏄🌊🏄🌊🏄

That's all for today, folks!

If you enjoyed this, please consider sharing this hand-crafted newsletter with a friend.