Harmonic raises $50 million to expand controls on workplace AI

Harmonic Security announced a $50 million Series B on October 14, 2026, led by Crosspoint Capital Partners with participation from existing investors. The company announcement distributed through Business Wire puts total funding at $76 million. The company says the money will support engineering and commercial expansion in the United States and United Kingdom.

Harmonic reports that its software covers millions of employees at more than 100 companies. Those are company figures. The financing announcement does not establish how reliably the product stops data exposure or how much time customers save.

What the software watches

Harmonic’s platform description centers on understanding the context of AI use. Its small language models classify what people are doing and the information involved. Security teams can choose to warn someone, block an interaction or record it for later review. The aim is to make restrictions follow a particular action rather than prohibit an entire application.

The agent security documentation extends that approach to tool calls. A local MCP gateway covers clients and servers on employee devices, while Harmonic Connectors cover supported cloud connectors in ChatGPT and Claude. Policies can distinguish reading a repository from deleting its contents, or allow a database query while refusing a write.

Harmonic says blocked calls receive an explanation that can help the agent try an allowed alternative. Invocation logs provide a record for investigation. These are documented product capabilities, not results from ByteForward testing.

Which features are available

The public package comparison separates three levels. Explore emphasizes discovery and usage analysis. Guide adds endpoint coverage and interventions. Command adds MCP discovery and permissions for agent actions. Buyers should establish which package covers the applications and connectors they actually use.

A notable limit is visible on that same page. Prompt injection detection and blocking is marked coming soon. It should not be treated as a shipping safeguard when evaluating the current product. The page lists package features without a public dollar price.

The monitoring decision matters too

Harmonic’s August 1 subscription agreement says monitoring may include personal accounts used through company managed devices, browsers or systems. It places responsibility for the observation scope and applicable employee notices and consultation on the customer. That is an implementation issue to settle before collecting workplace activity.

The agreement also says customer content is not used to train or modify model weights without prior written consent. Its security disclaimer does not promise to catch every exposure or policy violation, and describes the service as a supplement to a customer’s existing controls and judgment.

For a buyer, a useful evaluation would check both sides of a policy. Does a permitted task still finish, and does a prohibited action stop with a useful explanation? Reviewing the resulting logs also shows what information the organization would retain. Funding gives Harmonic room to expand, but those checks determine whether it fits a particular workplace.

This is a different job from the offensive testing workflow covered in ByteForward’s report on Armadin. One governs day to day AI interactions. The other probes an organization’s defenses for attack paths.

Illustrative network cable photograph by Leon Brooks, released into the public domain. Resized and converted to WebP. The photograph does not show Harmonic equipment.