Nexusflow raises $10.6 M, plans to build AI-powered cybersecurity interface

Nexusflow recently secured $10.6 million in a seed funding round led by Point72 Ventures, with participation from Fusion Fund and other respected players in the AI field at Silicon Valley. This funding round brings Nexusflow closer to its goal of delivering AI-powered cybersecurity solutions for enterprises.

Founded by UC Berkeley AI Research Lab Professors Jiantao Jiao and Kurt Keutzer and former ML Director at SambaNova Systems, Jian Zhang, Nexusflow aims to deliver a generative AI cybersecurity copilot fully powered by open-source models as an alternative to expensive, closed models such as GPT-4. Nexusflow Copilot boasts a 94% accuracy when function-calling from the CVE/CPE API, compared to GPT-4's reported 64% accuracy.

Broadly, Nexusflow Copilot should work as an assistant for enterprise security teams. Users can communicate with Nexusflow Copilot via a chat interface and ask it to perform several tasks using the enterprise's tools and knowledge resources. Nexusflow Copilot will then tap into the available resources to generate a candidate function that executes the task specified in the query. At the same time, Nexusflow Copilot learns about a particular workflow from the problems it is asked to solve by maintaining a past use corpus, so it better adapts to a specific style of use over time.

The magic behind Nexusflow Copilot is the newly released NexusRaven-13B, an open-source, commercially permissive LLM for function calling, comparable to ToolAlpaca or ToolLLM. Unlike some of these comparable LLMs, NexusRaven-13B generalizes on software it was not trained on. Generalization gives NexusRaven-13B a zero-shot function calling accuracy competitive with GPT-3.5. This model also has the benefit of being commercially permissive since it is not trained on GPT model generations. This means it can be applied for commercial uses competing with the GPT models without breaching OpenAI's terms of use policy. NexusRaven-13B was built and refined using the CodeLLaMA-13B, CodeLLaMA-34B-instruct, and LLaMA-70B-chat models.

While this is impressive, NexusRaven-13B's achievements aren't enough for an ambitious tool like Nexusflow Copilot. To achieve over 90% accuracy, Nexusflow developed a novel data curation methodology that involves feeding function data and context to CodeLLaMA-34B-instruct so it generates a description of the functions' capabilities. Then, LLaMA-70B-chat generates a natural language description of the code data and capability description. CodeLLaMA-34B-instruct also generates chain-of-thought traces that explain how to derive the values for the function arguments. Finally, each training data sample is enhanced with a list of similar functions to improve the function selection capability. This data curation procedure yields the desired results when paired with demonstration retrieval augmentation. A very detailed explanation is available on the Nexusflow blog.

For all the excitement, there is a bit of vagueness on the practical side of things: NexusRaven-13B was tested on the CVE/CPE and Virus Total APIs as well as software from the cybersecurity and generic domains, but it is not yet clear which specific apps and services Nexusflow Copilot supports. The company has expressed its plans to expand from single-round to multi-round interactions and continue to release commercially permissive models specialized in various other tasks. And even if a selling point of the open-source model is to give complete control and ownership of personalized models to the enterprises, the fact that Nexusflow Copilot seems to be a third-party add-on meant to integrate with the existent security solutions may not sit well with some clients once they consider the sensitivity of their responsibilities.