- Superhuman AI
- Posts
- Hunting for AI's security vulnerabilities
Hunting for AI's security vulnerabilities
ALSO: How to generate stunning AI images
Zain Kahn
June 26, 2024
Read time: under 4 minutes
Welcome back, Superhuman
Athletes already use AI to review past performances and make adjustments to their game plan. Now, startups are trying to apply that logic to the operating room — where even the smallest decisions can be a matter of life and death.
Today’s Insights
Breaking into AI systems to make them safer
Tutorial: Generate images with Freepik
Using AI to monitor operating rooms
5 new AI tools to boost your productivity
Everything else you should know today
AI-Generated Images: Poodle time
NEXT IN AI
AI companies are hiring hackers to sniff out security flaws
It's an AI startup’s worst nightmare: A hacker infiltrates its model, then uses it to generate dangerous content or extract sensitive data from users. Lately, though, tech companies have been paying hackers good money to do that very thing. These programmers use unusual prompting techniques to break through an LLM’s guardrails — with the ultimate goal of helping companies spot vulnerabilities.
The context:
LLMs have been finely tuned to block content related to sensitive topics — like instructions for how to concoct a dangerous drug
But there are different techniques to get around this, and some can be uncovered through simple trial and error
One involves feeding an LLM a long string of questions and answers to trick it into mimicking that behavior
These flaws also apply to image generators like Midjourney and DALL-E, which can spew explicit or violent material
What’s the solution? One new startup called Haize Labs partners with companies like Anthropic to purposely jailbreak models in order to expose their weaknesses. It’s developing a set of algorithms that can find security flaws and help AI companies fix them before they’re used for nefarious purposes. An Israeli security group called DeepKeep, meanwhile, helped Meta revamp its firewall after discovering a way to gain access to its users’ personal information.
But not everyone plays by the rules: Some vigilantes have taken matters into their own hands, complaining that AI companies aren’t acting fast enough. The Financial Times reported on one hacker known as Pliny the Prompter, who released a “Godmode” version of GPT-4o that’s stripped of Chat-GPT’s usual safety features. That version was taken down within hours, but experts say the incident proves that companies are struggling to keep up with the rate of prompting attacks.
PRESENTED BY HUBSPOT
Turn AI Into Your Personal Assistant
Discover how to turn AI into your personal productivity powerhouse with HubSpot’s highly anticipated AI Task Delegation Playbook. Master the art of AI delegation and optimize your workflow like never before. Get ready to save time and boost efficiency with their easy-to-use templates and calculators.
Leverage AI for streamlined task management, significantly enhancing time efficiency
Utilize AI tools to elevate decision-making and maximize workflow efficiency across teams
Explore comprehensive templates and detailed examples to master straightforward AI delegation
Evaluate and optimize productivity by assessing the measurable impacts of AI on your daily output
Don’t miss out—download your copy today and start transforming your workday.
AI AT WORK
How to generate stunning AI images using Freepik
Go to Freepik’s website and log in to your account. You will get credits to generate 20 AI images daily.
In the main menu, select the “AI image generator” option.
Describe your image in the prompt box and customize the image settings. You can change style, color, framing, and lighting.
After customizing click on Create button.
Wait for a few seconds and you’ll get multiple images based on your prompt.
PROMPT OF THE DAY
Cybersecurity for all
Prompt: Craft a detailed proposal emphasizing the importance of cybersecurity. Pitch this proposal to a [insert company here, for ex: hospital chain], highlighting the immense benefits of ensuring patient data security, and pinpointing the risks like [data breaches, patient privacy, and legal implications] that they might face without these services.
You can adapt the prompt to your specific needs.
Source: great ai prompts
PRESENTED BY GALILEO
Finally: Instant, accurate, low-cost GenAI evaluations
Why are Fortune 500 companies everywhere switching to Galileo Luna for enterprise GenAI evaluations?
97% cheaper, 11x faster, and 18% more accurate than GPT-3.5
No ground truth data set needed
Customizable for your specific evaluation requirements
AI & HEALTHCARE
AI could make surgeries safer — but not everyone’s onboard
Source: Surgical Safety Technologies
A 2016 Johns Hopkins study found that some 250,000 Americans die each year from medical mistakes, making them the third leading cause of death in the US. One reason: During surgeries, errors can be so subtle that there’s little opportunity to spot them — let alone learn from them.
When recording devices were added to planes in the ‘70s, accidents fell drastically. Now, AI startups are bringing that approach to the healthcare industry:
Surgical Safety Technologies (SST), founded by Stanford professor Teodor Grantcharov, uses cameras, microphones, and a tablet to collect around 500,000 data points per day
It then feeds the data through a specially-trained AI model that can highlight potential dangers
Even things like the proportion of off-topic conversations happening in the surgery room can reveal valuable clues about what went wrong
It’s a growing industry: Nvidia and Johnson & Johnson MedTech announced a partnership in March to collaborate on new AI-powered surgical devices, like a real-time tumor identification tool. Meanwhile, AI startup Theator helps surgeons rethink “their biases, suture choices, and how they think through [operations]” by giving them insights about which approaches are statistically safest. That’s important, experts say, because it’s sometimes the least-used methods that turn out to be most effective.
But not all surgeons are convinced: Some errors are unavoidable, especially during emergencies. And many surgeons are concerned that if their every movement is tracked, they could get slapped with lawsuits or other professional penalties. SST has a unique way of counteracting this problem. It anonymizes all of the data it collects, ensuring that no healthcare worker can be singled out. The focus is on “how we can build a system that makes it difficult for this to happen,” Grantcharov told the MIT Technology Review. “This is a system issue.”
PRODUCTIVITY
5 AI Tools to Supercharge Your Productivity
✅ Klue: Combine market, competitor, and buyer insights in a unified platform, driven by AI.
✅ ZebraCat: Transform your ideas into impactful marketing videos with AI voices, using a powerful text-to-video tool.
✅ MyLens: Use AI to create easy-to-understand visuals that highlight key insights and provide deep understanding.
✅ AI-FLOW: An open-source platform for creating custom AI tools through a simple drag-and-drop interface, designed for innovators and creators.
✅ LinkRobot: Boost your SEO with AI-powered internal linking, requiring no manual input from you.
PS: Want more? Check out our Top 100 AI Tools.
* indicates a promoted tool, if any
AI & TECH NEWS
Everything else you need to know today
Entering the Fray: Amazon is allegedly working on its own generative AI chatbot that will go head-to-head with ChatGPT. Code-named “Metis,” the platform could be released as soon as September.
Dream Team: OpenAI has purchased Multi, a startup that builds Zoom-integrated, collaborative video tools. The move comes just days after the ChatGPT developer acquired the database startup Rockset.
Plot Twist: Sources told Bloomberg that a potential deal between Apple and Meta was called off months ago due to safety concerns over the open-source nature of Meta’s AI offerings.
Beyond Books: A co-founder of the book-recommendation site Goodreads is back with a new platform called Smashing. It uses AI and community feedback to give recommendations across different mediums, including social media posts, articles, and YouTube videos.
😄 One Fun Thing: A platform called Inworld uses AI to make interactions with video game characters more realistic. To test the platform, the executive editor at the MIT Technology Review recently asked his colleagues to create a character based off of his own personality.
He came to regret the experiment: His virtual doppelgänger was “sarcastic to an extreme degree” with “an absolutely insufferable manner and — for some reason — an obsession with drinking the medieval beverage mead.”
AI-GENERATED IMAGES
Pick your Poodle
Source: @dianalouise on Midjourney
Prompt: a whimsical illustration of a tiny chocolate brown toy poodle with fluffy ears hanging down and a short tail with a pom-pom on the end, on a white background
Acquire new customers and drive revenue by partnering with us
Superhuman is the world’s biggest AI newsletter for businesses and professionals with 600,000+ readers working at the world’s leading startups and enterprises. Companies like Amazon, Hubspot, and Salesforce feature their products in Superhuman. You can learn more about partnering with us here.
