AI safety startup Haize Labs is raising early-stage funding that investors are eager to cash in on, according to information obtained by Business Insider.
The startup has received at least four term sheets from large venture firms in its early-stage funding round, according to two sources familiar with the matter, with crossover fund Coatue being one of the firms that has issued a term sheet, according to two sources familiar with the matter.
Haize Labs builds algorithms to test vulnerabilities in large language models and AI agents, a process known as AI red teaming, which has increasingly drawn investor interest, with a source telling Business Insider the startup has been a hot topic among investors over the last week.
Some startups have struggled to raise capital recently. But for hot startups, like Haize Labs' funding round, investor demand can be greater than the startup can fulfill. To land the deal, venture firms may try to compete on price. One source said Haize Labs received term sheets at valuations ranging from $30 million to more than $100 million.
Details of the round have not yet been finalized and the figures involved are subject to change. Haize Labs declined to comment.
The startup raised an undisclosed amount in a funding round earlier this year from Soma Capital, according to PitchBook. Other advisors and angel investors include the founders of Okta, Hugging Face, and Replit, as well as Google, Netflix, Stripe, and security operator Anduril, and professors from Harvard and Carnegie Mellon, Leonard Tan, one of the company's co-founders, told VentureBeat.
Haize Labs builds automated red teaming and stress testing algorithms to identify risks in AI models and agents. According to the company's website, Haize Labs aims to help customers determine guardrails for their systems, then “haze” them on an ongoing basis at scale to identify vulnerabilities and eliminate identified issues. The goal is to improve the robustness and reliability of these AI systems.
AI safety is becoming a growing concern, especially as AI usage increases for both consumers and businesses. What makes generative AI unique compared to traditional software systems is the ability of AI models and agents to create harmful content in the form of text, images, videos, etc. For example, OpenAI was previously accused of biased responses to user queries on ChatGPT.
In response, a wave of AI safety startups has sprung up, from Foreai, which evaluates LLM performance to improve accuracy, to Credo AI, which markets itself as an AI governance platform with safety guardrails for generative AI as one of its features.
Other startups like Haize Labs are applying cybersecurity principles to generative AI: Lakera AI invites users to insert malicious prompts to test the model for vulnerabilities, while Sama has a team of experts who uncover model weaknesses, and Adversa AI offers ongoing AI red team exercises on its underlying models.
In a viral post on X, Haize Labs revealed examples of vulnerabilities in AI models and open-sourced these “haize” on GitHub to highlight how easily the safety guardrails of major AI companies can be circumvented. Examples open-sourced include violent content in an image generation company and racist messages in a voice assistant.
The company was founded in December 2023 by three recent Harvard University graduates, Leonard Tang, Steve Li, and Richard Liu, Tang told VentureBeat. During their time at Harvard, the founders published 15 papers on machine learning, according to The Washington Post.
The company generates revenue through a services model to infrastructure providers and a SaaS approach for its automated “Haize” and defense solutions. Haize is already profitable and counts leading LLM developer Anthropic as a customer, Tang told VentureBeat.
