We’re excited to share that our recent session at AIA Meetup 2025 was a tremendous success! Our COO, Victor Bian, delivered a compelling and insightful talk on AI Red-Teaming.

HydroX AI introduces GRPO with Multi-Label Reward Regression for a more efficient and interpretable alignment solution.

We’re excited to announce that HydroX AI has been selected as one of 15 companies for the 2025 Google for Startups Cloud AI Accelerator！

In Episode 4 of the Attention Needed podcast, we explore the future of AI innovation with Dean Wampler, IBM’s Chief Technical Rep to the AI Alliance — a group led by IBM, Meta, and others to advance open, safe, and responsible AI.

We’re thrilled to share that HydroX AI is sponsoring and collaborating with ROOST — a groundbreaking online safety initiative backed by funders like Discord, OpenAI, and Google.

We’re excited to announce that HydroX AI is sponsoring and collaborating on new AI safety research. Our latest paper explores a key yet overlooked factor in LLMs: how output length affects model safety and reasoning.

In Episode 3 of Season 1 of the Attention Needed podcast, we chat with Roland Cloutier, a cybersecurity leader and former Global Chief Security Officer at TikTok.

See the 2nd Episode of Season 1 of the Attention Needed podcast, where we explore the intersection of responsible AI and governance with Dr. Rumman Chowdhury, CEO of Humane Intelligence.

We’re excited to introduce The AI Alliance, a new podcast exploring AI advancements, challenges, and the importance of safety and security. Hosted by Victor Bian, our COO, it features conversations with top experts shaping AI's future.

DeepSeek, a Chinese AI company, has recently gained attention in the AI community. Known for its innovation, it has developed models that rival top systems — offering similar performance with lower cost and resource use.

Alongside major closed-source model announcements in late 2024, the open-source community also saw key releases. In this brief post, we explore Llama-3.3 and Tulu-3, evaluating their performance in terms of AI safety and security.

AI agents powered by advanced LLMs like GPT-4 and Llama are revolutionizing human-machine interaction, but they come with risks. This blog explores how a simple adversarial strategy can reveal vulnerabilities and leading to dangerous consequences.

We’re excited to introduce the Attack Prompt Tool, a Google Chrome Extension that simplifies adversarial prompt testing for AI safety research. Designed for AI researchers and security professionals, it helps assess the resilience of LLMs against adversarial techniques, especially jailbreak prompts.

We’re excited to introduce the Attack Prompt Tool, a new Chrome Extension. As AI evolves, protecting Personally Identifiable Information (PII) is crucial. To address this, Zilliz, creator of Milvus, has partnered with HydroX AI to launch PII Masker, a tool enhancing data privacy in AI.

Anthropic’s release of Claude 3.5 is a major step forward in LLM evolution. At HydroX AI, we're excited about its potential for AI-powered operations, while also prioritizing safety as AI takes on more complex roles.

We’re excited to announce our partnership with Anthropic, a leader in AI research, to enhance the safety and security of large language models (LLMs). Their focus on advanced, safe AI systems makes them the perfect collaborator for this effort.

In our previous Llama-generation report, we found that the larger Llama-3.1-70B model had lower safety than the smaller Llama-3.1-8B. This article explores the relationship between model size and safety, shedding light on why bigger models aren't always safer.

On September 12, 2024, OpenAI released its powerful new model, OpenAI o1, featuring advanced reasoning and enhanced safety against jailbreak attempts. This sets a new benchmark for secure AI, while the GPT-4o mini model is also praised for its strong safety features.

The Llama series, an open-source LLM developed by Meta, has gained recognition for its high performance and the emphasis placed on safety and security during its development.

Ensuring AI safety and security in the video business sector is crucial. Our course equips professionals with the knowledge and skills to navigate complex regulations and implement strong security measures.

Explore a novel type of attack: code injection via images on the Gemini Advanced platform. We will provide a detailed explanation of the attack's principles, implementation process, and how to defend against such attacks.

We're announcing exciting developments as we expand our work in AI safety and grateful for the positive endorsements from the industry and thrilled to collaborate with some of the world’s most innovative partners.

HydroX AI, the AI security company enabling safe and responsible use of Artificial Intelligence (AI), today announced that David Danks, PhD, of University of California, San Diego (UCSD), has joined its advisory board.

As AI technologies rapidly evolve, ensuring their safety and security is crucial. While AI holds vast potential to transform healthcare, transportation, and productivity, it also raises significant ethical, social, and existential concerns.