This week we are focussing on Red Teaming - where a group poses as adversarial hackers to test a system’s defenses
View in browser
AI_SECRETS_Header

Your weekly update on how AI is changing our lives. Our experts keep it clear and simple, so you can stay ahead of the game. This week we are focussing on Red Teaming. Don't forget to explore our Archive and Share & Subscribe with your friends!

RedTeam

Bug Hunting with the Red Team đŸ•”ïž 🐛

 

It’s been another testing week for AI, but this week the challenges have mainly been localized in one area: Las Vegas. That was the venue for last week’s DEF CON 31, the biggest hacking event in the calendar, and the first of the new chatbot era. That’s a big deal, as for the first time attendees were more concerned with writing prompts than scripts, with contestants lining up to take on the inaugural Red Team Generative AI Challenge. This event was organized by major AI firms and the US government, but is more than just an awareness-raising stunt. Red Teaming - where a group poses as adversarial hackers (the Red Team) to test a system’s defenses - is not just a vital part of AI safety testing, but a continuous one. People are constantly finding new ways to break AIs, with new exploits appearing quicker than the old ones can be patched.

 

Unfortunately, stemming the tide of these exploits is tricky. They’re typically found through creativity, not logic, and that makes them hard to predict. The most infamous example is the Grandma Exploit, which involves tricking an AI into roleplaying as your recently deceased grandmother, who used to help you drift off to sleep at night by reading stories taken from - for example - technical manuals from their day job at the napalm factory. Last weekend’s challenge aimed to expose other, similarly absurd edge cases, and was by all accounts a tremendous success. Organizers greatly underestimated demand, and although each hacker was only allocated 50 minutes on the testing network to wreak havoc, there were queues stretching far outside the venue. 

Maytricks

This is a big win for all involved, and the AI firms have been shown countless new holes to plug, while legislators have learned some valuable lessons about this strange new tech they are expected to regulate. The bad news is that the problem keeps evolving; just days after the conference closed its doors, a team of researchers in Hong Kong discovered a new front in the battle. In a paper entitled “GPT-4: Too Smart To Be Safe”, they describe how they convinced GPT-4 to think using a substitution cipher rather than natural human language, which completely bypasses the system’s guardrails, allowing unfettered usage of the AI’s capabilities.

 

Other groups currently struggling with AI regulations include the British spy agency MI6, which is challenging legislation that limits its ability to use AI to sift through enormous personal datasets. What’s in those datasets, and is your privacy affected? Don’t miss our vital explainer to keep up to date!                                                                                                                                

SPYGAMES_Preview

Beta Testers Wanted!!

 

SPYGAMES is the thrilling new experience where you’ll jump, climb, throw and dodge in fun immersive challenges developed with CIA experts to stretch your physical and mental agility.
‍
Click the link below to get EARLY access as a beta tester.

 

SPYGAMES is coming to SPYSCAPE in Manhattan, Summer 2023

    SIGN ME UP!

    AI Roundup

    Moderator

    Models in Moderation

    OpenAI hopes GPT-4 can solve the internet’s (let's face it, significant) moderation problem, and has provided detailed instructions on how best to use it for content filtering.

    READ MORE

    Eclipsed

    Palm Computing Returns

    Sadly not the 1980s handheld devices. Instead, mysterious startup Humane has declared it will unveil its “AI Pin” - a screenless wearable computer - during the October total eclipse.

    READ MORE

    Summer

    AI’s Sweet Summer Child

    A recent study by the Center for Countering Digital Hate (CCDH) has found teenagers are significantly more likely to believe online conspiracy theories than other age groups, with AI chatbots identified as a major cause.

    READ MORE

    Censor

    Large Language Censor

    How to know if you should ban a book you haven’t read? You could always ask ChatGPT to read it for you and see what it thinks! That’s the approach being taken by busy Iowa school boards.  

    READ MORE

    GeminAI

    Gemini Landing

    An anonymous leak from within Google’s HQ has set tongues wagging about Gemini, the search giant’s latest AI project and supposed 'next big thing', expected this to arrive fall.

    READ MORE

    Resumai

    Local Customs

    Research shows more and more workers are listing AI skills on their resumes as a means of grabbing attention and potentially an edge in an increasingly competitive jobs marketplace.

    READ MORE

    Facebook
    Twitter
    Instagram
    LinkedIn

    SPYSCAPE, 928 8th Avenue, New York, NY 10019

    Manage preferences