Humane Intelligence and HydroX AI to host ‘Human vs. Machine’ event at RSAC 2025
“The increased availability of AI systems has led to greater interest in a wide range of applications—however, no technology should be blindly trusted,” said Dr. Rumman Chowdhury, ex-Twitter AI ethics lead. As artificial intelligence systems grow more powerful and accessible, questions around their security, reliability, and ethical use are taking center stage. At this year’s RSA Conference in San Francisco, a unique event aims to confront those concerns head-on — by pitting human intuition against machine logic. Humane Intelligence, a nonprofit known for its hands-on evaluations of AI’s societal impact, is teaming up with cybersecurity firm HydroX AI to host “Human vs. Machine,” an interactive red teaming session on April 28 at the Moscone Center. The two-hour event, scheduled from 1:10 to 3:10 p.m. Pacific time, invites participants to test the limits of large language models (LLMs) through 20 structured challenges. These will probe for vulnerabilities across topics such as hallucinations, political sensitivity, bomb threats, and prompt injections—issues that continue to spark global debate about the safety and ethics of generative AI. In technical terms, “red teaming” is a cybersecurity exercise that simulates real-world attacks to test an organization’s security with a goal to identify vulnerabilities and improve security. “Red teaming is one of the most effective ways to make AI safer. But it only works when we open up the process to diverse perspectives,” said Dr. Rumman Chowdhury, CEO and co-founder of Humane Intelligence. “This session is designed not just to test systems, but to teach people how to think adversarially and responsibly about AI.”
The event is part of RSAC’s Learning Labs, known for offering practical, hands-on training in emerging areas of tech and cybersecurity. Unlike traditional competitions, this red teaming session emphasizes collaborative learning over scoring points. Experts from Humane Intelligence, HydroX AI, and AIA will kick off the session with briefings on adversarial testing workflows and best practices in AI security. Victor Bian, COO of HydroX AI, emphasized the urgency of developing real-world skills. “We aim to make our RSAC Learning Lab a great opportunity for professionals to gain practical, hands-on skills in AI security,” said Bian. “By showcasing how humans and machines can work together to uncover vulnerabilities, we’re equipping attendees with actionable strategies to protect and strengthen AI systems.”
This session builds on Humane Intelligence’s growing leadership in the red teaming space, following their involvement in high-profile events like DEFCON 31’s generative AI testing challenge. As AI technologies continue to permeate every sector—from healthcare to finance—so too does the need for scalable and inclusive evaluation frameworks. Dr. Chowdhury, who previously held lead roles at Twitter as well as the U.S. government as the Science Envoy for AI, offered an exclusive comment to The American Bazaar, underscoring the importance of human judgment in an AI-driven world. “The increased availability of AI systems has led to greater interest in a wide range of applications—however, no technology should be blindly trusted,” she said. “The development of better test and evaluation systems, both AI and human-driven, can provide better insights into risks. Our event — ‘Human vs Machine’— tests the nuances of what AI can help automate, but where human oversight is still needed. While AI is helpful, contextual awareness can only be provided by people.”
As lawmakers, technologists, and civil society grapple with the pace of AI development, events like “Human vs. Machine” aim to equip stakeholders with the tools—and mindset—needed to secure an increasingly intelligent future.

Humane Intelligence and HydroX AI to host ‘Human vs. Machine’ event at RSAC 2025

Hessie Jones is a strategist, entrepreneur and investor covering AI
getty
New data shows enterprise AI adoption has jumped to 72%, up from 50% in previous years, according to McKinsey's 2024 Global Survey on AI but as companies rush to implement generative AI, they're hitting significant roadblocks. Most companies can set up generative AI systems in 1-4 months, and analysts expect that by 2025, AI will automate half of all digital work through an estimated 750 million AI-powered applications . But real-world results tell a different story. Gartner predicts 30% of generative AI projects will fail after initial testing by the end of 2025. "After last year's hype, executives are impatient to see returns on GenAI investments, yet organizations are struggling to prove and realize value," says Rita Sallam, Distinguished VP Analyst at Gartner. "As the scope of initiatives widen, the financial burden of developing and deploying GenAI models is increasingly felt." These problems show up clearly in practical use. Insurance companies using LLM products see just 22% accuracy with real business data . The success rate drops to nearly zero for complex tasks that need expert knowledge. This poor performance raises red flags about using the technology for essential business operations. The price tag for adding generative AI runs between $5 million and $20 million per organization. Yet some major companies are pushing ahead– JP Morgan recently gave AI assistants to 60,000 employees. This widespread rollout shows both possible cost savings and major companies' willingness to try generative AI despite its flaws. MORE FOR YOU
Experts say that while generative AI is new, it needs to be handled much differently handling than AI and machine learning systems. I recently met with Zhuo Li, former Director of Privacy and Data Protection at TikTok and current CEO of Hydrox.AI, which helps organizations implement security and compliance measures specifically designed for this new generation of AI, and David Danks, Professor of Data Science, Philosophy, & Policy at UC San Diego is also a member of the National AI Advisory Committee, and advisor to Hydrox.AI. We explore the mounting challenges organizations face with large language models. What is still unknown when it comes to evaluating the outcomes? What are the new attack vectors that can be created by LLMs; and what are the impacts when it comes to access to confidential or personal information, safety and society? The Shift From Single To Multi-Purpose AI
To distinguish generative AI from machine learning, Danks explains it this way: "The latest generation of AI systems that burst onto public consciousness with ChatGPT almost two years ago, are intended to be much more multi-purpose," Danks explains. "I want to use that term instead of general purpose. It isn't that they can do absolutely everything–I can't ask ChatGPT to make me an omelet, but I can ask it for a recipe for an omelet." "Since the early 2000s, the focus became, ‘let's find one task and build a system that is optimized for that one task and far exceeds human performance,’” says Danks. "Whether it's predicting sepsis in intensive care unit patients or doing image classification as a picture of a dog versus a picture of a cat." Another major change lies in how users interact with these systems. Current AI interfaces have evolved to use natural language prompts, allowing direct questions and responses between users and large language models. This shift has made AI systems more interactive and user-centric, as individuals can tailor the AI's output to their specific needs and interests. As Danks points out, the interface evolution is a significant departure from earlier systems that required more structured or technical input methods. However, he points out an important trend in practical applications: "Over the last two years, we're increasingly seeing people move away from using what we might think of as the base version of a generative AI system. We're increasingly seeing people move towards what we often refer to as fine-tuned AI systems." These specialized versions serve specific purposes while maintaining their language-centered, multi-purpose nature. As Danks explains, "There does still seem to be a certain amount of generality versus performance tradeoff. If we want to increase performance, we still need to give the system special information." This shapes how businesses implement AI. "I don't want people to think that we are just constantly expanding the scope of what we can do with an AI system," Danks cautions. "Because in practice, we're oftentimes narrowing it back down for particular use cases we have, especially in the business context." New Accessibility And 'Monoculture' Risks in LLMs
The fundamental problems that have plagued AI for decades remain present in current systems. "Generative AI, foundation models exhibit really impressive performance, but they still get things wrong," Danks explains. He points to what many call "hallucinations" - though he views these as simply traditional AI errors in new forms. A significant new concern stems from how easily accessible these systems have become. Danks contrasts this with earlier approaches: "Historically, when we would build AI systems... there would be an intensive training process with the user to help them understand. Here's how you use it. Here's the kinds of things that are good inputs versus bad inputs." Today's prompt-based interfaces, while user-friendly, have eliminated these safeguards. This accessibility has led to both intentional misuse and accidental errors. "People doing things maliciously, trying to break the system... but it's also led to a number of accidental errors or accidental maliciousness," Danks notes. Perhaps most concerning is what Danks terms "algorithmic monoculture." This risk emerges from the concentration of AI development among a few major companies– Anthropic, Meta, OpenAI, and Google–all training their models on similar datasets scraped from the web. "These models are massive. They take an enormous amount of compute power, vast amounts of data, and an immense amount of processing ability... so there really are only a few companies in the world that are able to create these kinds of systems," Danks explains. The risk isn't simply about the limited number of foundation models but he points to the subtle danger: "They're being trained on largely a lot of the same data because they're all scraping as much data as they can off of the web. And as a result, we're seeing the convergence of some of these generative AI systems towards all being similar." This convergence creates a significant security vulnerability. "If everybody's using exactly the same system, then any shortcoming or failure in that system is going to have widespread effects," Danks warns. While he acknowledges that monocultures aren't inherently problematic–citing how most internet servers run similar code–he emphasizes a crucial distinction: "It's OK because we put a lot of energy and effort into safety and security, however, we aren't yet seeing the same level of focus on safety and security for generative AI systems." Critical Security Gaps In Generative AI Systems
According to Zhuo Li, CEO of Hydrox.AI, large language models present three fundamental security challenges that set them apart from traditional systems. Traditional cybersecurity primarily focuses on access control. This involves preventing unauthorized users from gaining system privileges or accessing confidential information in databases. Most people have a basic grasp of these concepts and understand the implications of a lost password or someone gaining unauthorized permissions. However, when it comes to generative AI systems, the situation is quite different. He explains how large language models process text:
When you interact with an AI system like ChatGPT or Claude, here's what happens behind the scenes:
First, imagine a massive program (what engineers call a "binary") running on powerful computers somewhere in the cloud. This program contains the AI model that processes your request. When you type a simple question (called a prompt) like ‘Where can I buy candy?’, the system begins a series of steps:
1. Your question first gets broken down into smaller pieces called tokens. These might be words, parts of words, or even individual characters. 2. These tokens are then converted into numbers through a process called embedding. Think of this as translating your human language into a mathematical form the AI can understand. 3. These mathematical representations then travel through what Li describes as a vast network of interconnected nodes, each one helping to process your question and figure out the most appropriate response. 4. At each step, the system uses probability calculations to determine what word or token should come next in its response. This is why you have seen AI responses appear one word at a time - the system is actually generating each piece sequentially, constantly calculating what should come next. This process highlights why AI systems can be unpredictable–they are making probability-based decisions at each step, which means asking the same question twice might give you slightly different answers each time. Li demonstrates how this complexity justifies why security approaches don't work well with these systems. Unlike a traditional computer program that follows fixed rules, these AI models make countless probabilistic decisions in ways that can be difficult to predict or control. Li points out, “But how do you secure a system that makes thousands of probability-based decisions for every response? How do you interpret the output, and more importantly, how much trust can we place in it?”
Understanding interpretability as well as the control is still nascent and Hydrox is actively working on these challenges through evaluation, continuous monitoring, and comprehensive security testing. In addition, they apply Red Teaming , a “comprehensive, adversarial approach to testing an organization's security posture by simulating real-world attacks” in identifying AI system vulnerabilities. According to Li, this testing focuses on two distinct types of risks. First, "default issues" –inherent problems within the model that can surface during normal use, where the AI might generate inappropriate or problematic responses even without any malicious intent. Second involves active attack scenarios, where security teams deliberately attempt to manipulate the system. As Li explains, these tests involve trying various "tricks" to see how the model behaves under aggressive probing. Such attacks can have serious consequences, potentially leading to the generation of inappropriate content, exposure of sensitive information stored within the model's training data, circumvention of built-in security measures, or even breaches of personal data. A New Frontier In Corporate Espionage
Large language models are opening unprecedented avenues for extracting confidential information, transforming from potential targets of attacks into powerful tools for corporate intelligence gathering, according Danks. "These systems are incredibly good at synthesizing very large amounts of data, so I can give a large language model 1,000,000 pages of text, and they can very quickly synthesize it and draw patterns and insights..." This capability creates a new kind of security threat that many organizations haven't yet recognized. The risk emerges from companies' routine public disclosures. "A lot of companies and organizations will have publicly available documents in patent filings or where some proprietary information is perhaps being leaked through," Danks notes. While this information has always been technically accessible, its sheer volume has made it impractical for human analysis. That protection through obscurity is disappearing. As Danks describes, "Now I can go to my competitor and I can scrape everything they've ever made publicly available and ask a large language model to synthesize to extract what are the underlying technology that would explain the different patents they hold or where should I look to find out more information about what's going on at this company." The threat becomes even more sophisticated when considering LLM-versus-LLM scenarios. An attacker could use one LLM to analyze and probe another LLM that has been customized with a company's public documentation. This process can uncover patterns and insights in minutes that might take human analysts years to discover. "Previously, we could make certain things publicly available, and we didn't have to worry about privacy violations because no one’s going to go through all of that effort," Danks explains. "Well, that's all changed with large language models." Balancing AI Innovation With Protection
How do we establish trust in technology that's still proving itself? "The vast majority of companies that I know are starting to use generative AI systems are basically saying if something is mission critical, do not assume that it's right," Danks explains. "Make sure that a human is actually checking it over." "We're seeing the very rapid emergence of a number of what I would call governance mechanisms," Danks explains. "Regulation is one, but there are many others– voluntary standards, safety and security testing that companies are developing, and efforts to establish best practices similar to what we see in cybersecurity." However, we haven't reached a definitive turning point in AI governance yet. "These systems have really only been in the wild, in the public for under two years," he points out. Instead, what's emerging is what Danks calls "a kind of iterative loop" - trying new governance approaches, seeing what works, and adjusting accordingly. Sometimes this means updating best practices; in other cases, it might mean recognizing the need for stronger regulation "because industry is not going to do it on their own." The global nature of AI technology adds another layer of complexity. "We're right now seeing radically different governance ideas, regulatory ideas show up in the US versus the EU," Danks explains. While Europe pursues broad AI regulation through the AI Act, the US focuses on sector-specific rules. Meanwhile, China takes yet another approach entirely. As Danks observes, "Companies, in a certain sense, don't respect national borders - they all want to operate worldwide." This raises a crucial question: "How do we start to make the different regulatory governance regimes that are being stood up... consistent with one another so that companies know what to plan for, what to design for, what the standards are?" He describes the current state of AI governance as a roller coaster, with an uncertainty of where it eventually lands. Trust and technology have always been at odds, but governance is bringing us closer to minimizing this gap, as per Danks, "We're entering the creative, innovative, iterative phase of governance and regulation right now. And on the one hand that's very exciting because it means there's a lot of possibilities... but at the same time, it's kind of a terrifying phase to be in because we know we're going to make some mistakes." Follow me on  Twitter  or  LinkedIn . Check out my  website  or some of my other work  here .

The Hidden Security Costs Of Rapid Generative AI Implementation

HydroX AI is working with big-name tech firms and the AI Alliance to evaluate generative AI models in industries such as healthcare and finance. HydroX AI, a startup developing tools to secure AI models and services, is teaming with Meta and IBM to evaluate generative AI models deployed in high-risk industries. Founded in 2023, the San Jose, California-based company built an evaluation platform that lets businesses test their language models to determine their safety and security. HydroX will work with Meta and IBM to evaluate language models across sectors including health care, financial services and legal. The trio will work to create benchmark tests and toolsets to help business developers ensure their language models are safe and compliant before being used in industry-specific deployments. “Each domain presents unique challenges and requirements, including the need for precision/safety, adherence to strict regulatory standards, and ethical considerations,” said Victor Bian HydroX’s chief of staff. “Evaluating large language models within these contexts ensures they are safe, effective, and ethical for domain-specific applications, ultimately fostering trust and facilitating broader adoption in industries where errors can have significant consequences.”
Benchmarks and related tools are designed to evaluate the performance of a language model, providing model owners with an assessment of their model's outputs on specific tasks. The startup is now working with two major tech companies that have experience working on AI safety. AI Focus
Meta previously built Purple Llama, a suite of tools designed to ensure its Llama line of AI models is deployed securely. IBM meanwhile was among the tech companies that pledged to publish the safety measures they’ve taken when developing foundation models at the recent AI Safety Summit in Korea. Meta and IBM were founding members of the  AI Alliance , an industry group looking to foster responsible open AI research. HydroX has also joined and will contribute its evaluation resources while working alongside other member organizations. “Through our work and conversations with the rest of the industry, we recognize that addressing AI safety and security concerns is a complex challenge while collaboration is key in unlocking the true power of this transformative technology,” Bian said. “ It is a proud moment for all of us at HydroX AI and we are hyper-energized for what is to come.”

Meta, IBM Working With Startup to Test AI Model Safety

HydroX AI ( see Pulse 2.0 profile here ), an AI-based company with the industry’s first comprehensive all-in-one large language model (LLM) safety and security platform, announced the closing of a $4 million angel round from Vitalbridge Capital, Atom Capital, and AI expert Qi Lu. This funding will provide the means for HydroX AI to continue LLM safety innovation, validate demand, and further the platform’s market entry goals. And the funding comes during a fast-growing segment of the exploding LLM AI market. And global LLM revenue is projected to grow to 260 billion by 2030. HydroX AI has found that a typical LLM is vulnerable to attack and compromise in as little as three months from release, which means constant security measure updates are a must. So HydroX AI sees great need and potential in the AI Trust, Risk, and Security Management (AI TRiSM) sector. HydroX AI was created to address these LLM risks by offering a comprehensive platform for the evaluation, mitigation, protection, and monitoring of Large Language Model (LLM) safety and security in the four trust domains: safety, privacy, integrity and security. HydroX also announced the official launch of EPASS, which is its cutting-edge Evaluation Platform for AI Safety and Security. And this platform enables precision evaluation and management of AI models, offering detailed insights that empower users to understand and enhance the safety, security, privacy, and integrity of their AI systems. Plus, EPASS serves as a crucial tool in a landscape where robust evaluation frameworks are imperative due to the substantial ethical, social, and existential risks associated with AI technologies. The platform enables stakeholders across industries to manage and compare AI models easily, enhancing decision-making and improving AI system reliability without having to build an expensive in-house evaluation team and infrastructure. All standard accounts come with essential features, including one model evaluation daily, one attack method, and one evaluation per category. HydroX AI is now in early development, engaging with major LLM companies. Several of these companies have seen notable enhancements from the platform, showcasing EPASS’s potential and affirming its strategy for tackling trust issues in LLMs. These endorsements from leading LLM companies highlight its pioneering role and also secure a first-mover advantage in the crucial, evolving field of AI safety and security. HydroX AI was launched by Zhuo Li – who was the former head of the Privacy and Data Protection Office at Bytedance. Before Bytedance, Li developed his expertise at Meta and LinkedIn, contributing significantly to privacy and security measures. And the team includes Head of Engineering Yuji Kosuga, with deep roots in the field through experiences at companies such as LinkedIn, Google, and Meta, underscoring their profound dedication to advancing AI safety and security. KEY QUOTE:
“The blending of artificial intelligence with society is just starting. It raises issues beyond privacy and security, including ethics, learning, and mental effects. Looking ahead, security will be key, not just for protection but as a foundation for AI’s growth.“
– Ben, partner at Vitalbridge Capital
Pulse 2.0 focuses on business news, profiles, and deal flow coverage.

HydroX AI: Trust, Risk, And Security Management Company Raises $4 Million

News provided by
Share this article
The AI Startup Aims to Deliver Safety and Security for AI Large Language Models
SAN JOSE, Calif., June 25, 2024 /PRNewswire/ -- HydroX AI, an artificial intelligence (AI) startup with the industry's first comprehensive all-in-one large language model (LLM) safety and security platform, today announced closing a $4 million angel round from Vitalbridge Capital, Atom Capital and noted AI expert, Qi Lu, EVP of Microsoft, former COO of Baidu. The funding will provide the means for HydroX AI to continue LLM safety innovation, validate the market need, and further the platform's market entry goals. HydroX AI EPASS Platform
The investment comes to a fast-growing segment of the exploding LLM AI market. Global LLM revenue is projected to grow to 260 billion by 2030. With that growth comes increasing safety and security risk. HydroX AI has found that a typical LLM is vulnerable to attack and compromise in as little as three months from release, meaning constant security measure updates are a must. As a result, HydroX AI sees great need and potential in the AI Trust, Risk, and Security Management (AI TRiSM) sector. "The blending of artificial intelligence with society is just starting. It raises issues beyond privacy and security, including ethics, learning, and mental effects. Looking ahead, security will be key, not just for protection but as a foundation for AI's growth," said Ben, partner at Vitalbridge Capital. HydroX AI was created to directly address these LLM risks by providing a comprehensive platform for the evaluation, mitigation, protection, and monitoring of Large Language Model (LLM) safety and security in the four trust domains: safety, privacy, integrity and security. Introducing EPASS - HydroX AI's Evaluation Platform for AI Safety and Security
HydroX also announces the official launch of EPASS , its cutting-edge Evaluation Platform for AI Safety and Security. HydroX AI EPASS Platform
This state-of-the-art platform enables precision evaluation and management of AI models, offering detailed insights that empower users to understand and enhance the safety, security, privacy, and integrity of their AI systems. EPASS serves as a crucial tool in a landscape where robust evaluation frameworks are imperative due to the substantial ethical, social, and existential risks associated with AI technologies. The EPASS platform facilitates:
Evaluation: Precision assessment of AI models using advanced techniques to identify vulnerabilities and ensure models meet high standards of safety and integrity. Detailed reports provide actionable insights, helping users refine AI performance and reliability. Management: A user-friendly dashboard allows for effortless model management and comparison, integrating seamlessly into existing workflows for real-time monitoring and updates. Leaderboard: A comprehensive ranking system that benchmarks AI models across over 30 safety and security categories, offering users critical reference points for continuous improvement. This platform allows stakeholders across industries to manage and compare AI models with ease, enhancing decision-making and improving AI system reliability, without the need to build an expensive in-house evaluation team and infrastructure. All standard accounts come with essential features, including one model evaluation daily, one attack method, and one evaluation per category. For those seeking advanced and comprehensive evaluations, HydroX also offers a range of customizable add-ons:
Unlimited Evaluations per Category: Dive deeper into the evaluation process with unlimited assessments for each category. Category Bundles: Tailored category bundles designed to align with specific domains, streamlining the evaluation process. Higher Speed: Accelerate evaluations with higher processing speeds, optimizing efficiency. Dashboard to Visualize Results: Gain real-time insights into models' performance with intuitive dashboard interface. More Attack Methods: Expand assessment capabilities with a wider array of attack methods at the clients' disposal. Dedicated GPUs: Enhance processing power and efficiency with dedicated GPU resources. HydroX AI is currently in early development engaging with major LLM companies. Several of these companies have seen notable enhancements from the platform, showcasing EPASS's potential and affirming its strategy for tackling trust issues in LLMs. These endorsements from leading LLM companies not only underline its pioneering role but also secure a first-mover advantage in the crucial, evolving field of AI safety and security. HydroX AI was founded by Zhuo Li, former head of the Privacy and Data Protection Office at Bytedance. Before Bytedance, Li developed his expertise at Meta and LinkedIn, contributing significantly to privacy and security measures. The team, rich in safety and security veterans, includes Head of Engineering Yuji Kosuga, with deep roots in the field through experiences at companies such as LinkedIn, Google, and Meta, underscoring their profound dedication to advancing AI safety and security. About HydroX AI LLC
Founded in 2023, HydroX AI ( www.hydrox.ai ) is a San Jose, California based AI trust, risk and security management (AI TRiSM) startup with the mission to ensure a safer world through cutting-edge AI security solutions for LLM evaluation and enhancement. Their platform, the industry's first all-in-one platform for LLM safety and security, allows LLM developers to address the major LLM risk areas with development-stage tools and post deployment monitoring tools. Check out https://www.hydrox.ai/blogs/64ab58a5-73a9-4b82-bffa-dcb2406a3efe , for a more in-depth look at how to use EPASS. Additionally, dive into our newest blog post highlighting a white paper exploring the vulnerabilities of LLMs, dissecting attack methods, and offering practical recommendations for anyone interested in AI security: https://www.hydrox.ai/blogs/818992e9-b828-45c3-b808-b69dca5a5833 . SOURCE HydroX AI

HydroX AI Announces Closing of $4 Million Angel Funding Round and Launches EPASS

ByteDance

LinkedIn

Hewlett Packard Enterprise

VMWare Technology Alliance Partner Program

VMware

Microsoft

7 tech M&A predictions for 2025

These companies protect organizations from digital threats.

Cybersecurity

Companies working on generative AI applications and infrastructure. 

Generative AI

Artificial Intelligence

Incubator/Accelerator

HydroX AI engages in the development and commercialization of artificial intelligence (AI)-assisted safety solutions. The company offers language model (LLM)-supported risk assessment services, ethical AI frameworks, cybersecurity solutions, and more. HydroX AI primarily caters to businesses, government bodies, and cybersecurity and ethical hacking professionals. It was founded in 2023 and is based in San Jose, California.

Name	Zhuo Li
Work History	ByteDance, Meta, LinkedIn, Hewlett Packard Enterprise, VMWare Technology Alliance Partner Program, VMware, and Microsoft
Title	Founder, Chief Executive Officer
Status	Current

How VCs Use CB Insights

Professional Services

Platform Overview

HydroX AI

Executives

HydroX AI Management Team

1 Team Member