In partnership with

Dear Sentinels

AI systems have snuck into our daily software toolkit, whether we notice or not. From chatty customer support bots to coding assistants and those autonomous agents that merrily send emails, query databases, and even do a bit of shopping for us, they're everywhere. Oddly enough, most of the security chatter about these systems gets stuck on the usual suspects like jailbreaks and hallucinations, while the rest of the attack surface is left to gather dust. This is where threat modelling comes in handy: it lets us take a step back and ask, in a systematic way, what could possibly go wrong. STRIDE, that trusty framework from Microsoft in the late 1990s, is still one of the go-to tools for this sort of thing.

STRIDE splits threats into six neat categories: spoofing, tampering, repudiation, information disclosure, denial of service, and elevation of privilege. Once you start poking around AI systems, you’ll see each of these fits right in. For example, an attacker could pretend to be a trusted tool or user, slip something nasty into the training data, or sneak hidden instructions into a document the model fetches. Models can also blurt out sensitive information, and if an agent starts taking actions no one can trace, you’ve got yourself a repudiation headache. Prompts that gobble up resources can lock everyone else out, and perhaps the most hair-raising is elevation of privilege: if you can sweet-talk a model into using tools it shouldn’t, you’ve just turned a text-based exploit into a real-world mess. In this edition, we’ll stroll through each category, see how it plays out in AI, and offer some practical tips for keeping things in check.

Last week, we took STRIDE for a spin with IoT. This week, we’re zooming out to tackle the grander world of AI. If someone ever decides to mash those two together, well, brace yourself.

Read less, know more.

That's it. That’s the pitch.

Morning Brew is a free daily email that explains the news shaping your life in a way you won't hate. From how geopolitics hits your wallet to AI eating entire industries, it's all explained in about 5 minutes, with just enough puns to keep things interesting.

Heck, you might even enjoy the news for once. No jargon. No pretending you understood that stock market headline.

See why over 4 million people start their day with Morning Brew.

News from around The Web

Subscribe to keep reading

This content is free, but you must be subscribed to AI Sentinel Newsletter | Exploring the frontiers of AI Research and Cybersecurity to continue reading.

Already a subscriber?Sign in.Not now