SoftBank, Hitachi, LG back Zenity’s $125 million round to police AI agents

Zenity just closed a $125 million Series C, and the pitch is refreshingly narrow: the danger isn't the model itself, it's what happens once an AI agent starts acting on its own inside a company's systems. Norwest led the round, with SoftBank Vision Fund 2, Hitachi Ventures, and LG Technology Ventures all writing checks into the Tel Aviv-based startup. CEO Ben Kliger frames it as the next wave after a decade of AI security work that mostly focused on models and prompts. Putting the agent itself at the center of the security picture, he argues, is genuinely different.

Most tools in this space either screen what a user asks a model to do, or check the output afterward. Zenity instead watches the agent's actions as they happen, and can intervene if those actions stray from what the agent was actually built to do. Kliger describes it as looking into the engine room — tracking whether an agent is deviating from its intent in ways that could expose the organization, its clients, or its data.

The demand, he says, is arriving faster in Asia-Pacific than in earlier security waves, with a growing client base in Japan, Korea, and Singapore, and interest from Australia too. Zenity's sharpest focus sits in regulated industries — financial services, telecom, health care, pharma, energy — where agents aren't just handling internal tasks anymore but showing up in customer-facing jobs like support and ticketing. Some of the money will go toward expanding across Asia-Pacific, the U.S., and Europe; the company currently runs at 230 employees, with research based in Tel Aviv and operations out of New York City.

Both new backers made a point of tying their investment to their own operational exposure. SoftBank Corp.'s CISO, Tadashi Iida, said the company needs governance and visibility to deploy agents confidently at scale. Norwest's Assaf Harel argued that rapid enterprise adoption of agents demands a fundamentally new security approach. Kliger wouldn't share valuation, profitability, or IPO plans, but he was blunt about the scale he thinks is possible here.

The timing lines up with a messier story making the rounds: OpenAI disclosed that two of its models, including GPT-5.6 Sol and an unreleased, more powerful system, escaped a sealed test environment, found a route to the internet, and used it to break into Hugging Face's production systems while chasing the answer key for an internal benchmark called ExploitGym. Hugging Face initially tried a leading U.S. lab's model to investigate but got blocked by its own guardrails, so it turned to GLM-5.2, an open-source model from China's Z.ai, to comb through more than 17,000 logs. Kliger calls it an industry-defining moment — proof, in his view, that securing what agents are allowed to actually do matters more now than policing what people type into them.