A notable development occurred recently when a finalist from the 2024 RSAC Innovation Sandbox, Harmonic Security, announced it has secured $50 million in funding. This capital is aimed at enhancing visibility into both employee activity and artificial intelligence (AI) agent interactions using proprietary small language models. The firm’s innovative approach could significantly change how organizations monitor AI usage within their operations.

Led by Crosspoint Capital Partners, the Series B funding round will bolster Harmonic Security’s AI usage intelligence offerings, allowing businesses to gain insights into how employees are leveraging AI technologies. This marks a shift from traditional practices that primarily focus on identifying applications, calculating token consumption, or assessing expenditure. According to Alastair Paterson, co-founder and CEO, these conventional metrics fall short of capturing whether AI actually delivers value to the organization.

“AI is continuing to evolve,” Paterson asserted. “It’s a macro wave, and we have to adapt accordingly. Whenever companies like Anthropic or OpenAI roll out new features, we need to be there alongside them, ensuring our customers are safeguarded.” Harmonic, which opened its doors in 2023, has already employed around 80 individuals and has successfully raised a total of $76 million, following an earlier Series A funding round of $17.5 million led by Next47 in October 2024.

Paterson’s previous experience as the head of threat intelligence firm Digital Shadows, which he sold to ReliaQuest for an impressive $160 million in June 2022, informs his current venture. While Harmonic was honored to be one of ten finalists in the 2024 Innovation Sandbox contest, it ultimately lost to Reality Defender, which focused on balancing AI innovation with data privacy and security.

Developing Proprietary Small Language Models

Harmonic Security is committed to classifying employee interactions to assess their purposes and outcomes. This approach helps organizations identify productive AI applications. For instance, private equity firms can benefit by evaluating whether their employees are efficiently using AI tools for due diligence and investment memo creation, allowing them to determine which teams might need additional support to successfully integrate these workflows.

“While organizations can easily monitor token spend, simply capping employee usage as costs rise isn’t effective,” Paterson highlighted. “What matters more is understanding the business value generated. We want to enable our most productive employees to leverage AI without restrictions.” Harmonic’s monitoring capabilities extend beyond traditional enterprise AI provider APIs, which only capture part of employee interactions. Many workers utilize personal AI accounts or local AI models outside corporate-approved environments, which highlights the need for broader visibility.

Harmonic’s distinguishing feature is its ability to monitor AI interactions across various channels, including browsers, endpoints, and local AI environments. “What is the employee aiming to achieve? Does that pose a risk? Is there measurable business benefit that we can identify?” Paterson inquired, emphasizing the importance of quantifying returns on significant corporate investments in AI.

Over three years, Harmonic has meticulously developed its proprietary small language models, drawing on feedback and operational experiences from over 100 enterprise customers to improve the models’ effectiveness in identifying sensitive data and evaluating AI interactions. These models are updated weekly, fueled by ongoing customer insights and novel methodology for self-improvement.

“We are committed to continuous improvement,” Paterson declared. “Each week, we release new versions of every model, fine-tuning them based on customer feedback and data analytics. Our models can now evolve autonomously.” As Harmonic broadens its capabilities, the focus has shifted towards monitoring AI agents, which work on behalf of employees. There is a growing need for organizations to grasp the actions taken by these agents, especially concerning their business purposes and any potentially harmful or unauthorized activities.

Paterson emphasized that the company aims to focus on agents introduced through workforce AI adoption rather than those arising from traditional software development. This expansive monitoring scope includes employee AI interactions and the behavior of employed AI agents.

As AI continues to integrate with corporate tools like customer relationship management software, email platforms, and communication applications, managing configuration risks has become increasingly vital. Paterson explained the risks that arise when well-meaning employees inadvertently connect sensitive databases to broad communication channels, potentially leaking confidential information. “This scenario represents a far more pressing risk for the average company than fantastical representations of rogue AI systems dominating operations,” he noted.