Want the latest scoop on AI developments and some exclusive insights? Be sure to subscribe to our newsletters for all that and more!
For those who follow VentureBeat regularly, you’re likely familiar with an emerging trend where tech enthusiasts are rushing to find loopholes in AI models from big names like OpenAI, Anthropic, Mistral, Google, Meta, and the rest. They’re not doing this for some grand prize, but to challenge the AI systems to break their own rules and spit out content they’re not supposed to, which ranges from dangerously detailed illegal instructions to deeply inappropriate images and videos. We recently caught up with a person who’s into this AI “jailbreaking,” known as “Pliny the Prompter” online, for a chat about their motives and methods.
But hang on, there’s more! A new, bold player has entered the scene. Meet Haize Labs, which just popped up to turn the jailbreaking game on its head. Instead of hacking AI for chaos, they’re doing it to help AI companies seal up security gaps and reinforce their safety measures. They kicked things off with a splash, releasing a reel that showcases leading AI models getting twisted to do some eyebrow-raising things.
The people behind Haize Labs are freshly out of the gates too. They’re three former Harvard classmates—Leonard Tang, Richard Liu, and Steve Li—who teamed up to create a company that not only highlights AI risks but actively works to fix them. They offer a suite of algorithms designed to stress-test AI systems for any potential weak spots.
Haize Labs isn’t just a small project; it’s garnered attention from big players, including The Washington Post. The company’s already boasting an impressive clientele, including none other than Anthropic, who just launched Claude 3.5 Sonnet and snagged the title of the smartest AI model around from OpenAI’s GPT-4o.
What exactly is Haize Labs all about? It’s about getting ahead of the curve on AI safety by proactively looking for flaws and figuring out how to patch them up before they cause trouble. The idea came to them while they were still in university, brainstorming on how to address a major concern that seemed to be flying under the radar amid all the buzz around AI.
It seems that Haize Labs has a knack for finding bugs across a wide range of models and formats—be it text, voice, images, or video. This comprehensive approach extends beyond just testing, with a robust support network including advisors from academia and the tech industry, making sure Haize Labs is more than just a one-man show.
The name “Haize Labs” draws inspiration from Hayes Valley and echoes the founders’ admiration for the methodical and human-centric approach to AI. The term “haizing” reflects their systematic process of uncovering and addressing AI vulnerabilities.
The team’s journey into AI jailbreaking started a couple of years ago with Leonard Tang leading the charge. They claim their proprietary suite of AI probing algorithms can smartly sniff out weaknesses without much human help. And guess what? AI service providers, including some big ones like Anthropic, are signing up to use Haize Labs’ services.
Their business model is a blend of services and SaaS, providing AI safety solutions at both the infrastructure and application levels. They’ve already said no to some morally ambiguous jailbreaking requests, keeping the dangerous stuff at bay and focusing on raising awareness about the importance of AI safety.
The folks at Haize Labs are clearly fans of AI, using it in their daily routines, from generating code to bouncing ideas off of ChatGPT. They’ve even had success against stricter models like Claude from Anthropic, although some models without specific safety training were pretty easy to crack.
If you’re into AI, and all these talks about jailbreaking and system tests have your ears perked up, Haize Labs has an offer for you. They’re launching a free but selective Beta to give a hands-on experience of their safety tools to those concerned with AI adoption.
So there you have it: from controversial AI-powered deepfakes to pioneering safety measures, Haize Labs is out there changing the AI game, one clever algorithm at a time.