If you're building an AI system that makes decisions or takes automated actions, clients will start asking how you prevent unexpected behaviour. Prepare a specific technical answer: which steps you log, where you have stop points, and what happens when the system tries to do something outside scope. Large organisations are already paying for a separate oversight layer, and this will reach smaller projects too.
What happened
Onyx Security, an Israeli company founded in 2024, closed a $113 million Series B funding round led by Bessemer Venture Partners. This happened just four months after the company emerged from stealth with $40 million. The company has now raised $153 million in total since its founding.
The company built a platform that lets organisations discover, manage and control the use of AI agents – both those developed internally and third-party ones. The platform monitors every step in an agent’s reasoning process, identifies malicious or unintended behaviour, and intervenes to prevent harmful actions before they occur. The company employs over 80 people in Israel, the US and Canada, and works with many Fortune 500 companies. In June, Anthropic announced it had integrated Onyx technology.
Why it matters beyond the headline
This funding marks a market shift: large organisations are willing to pay significant amounts for a dedicated security layer for AI agents. This is not another general security product – it’s a platform focused on one specific problem: what happens when an AI system makes decisions or takes actions on its own.
The fact that Anthropic, one of the world’s leading model companies, uses Onyx technology indicates global relevance. This is not a solution built only for the local market. Israeli companies building complex AI systems should know Onyx as a potential partner or technical competitor – it sets the local standard for AI governance.
The speed of the funding – four months between rounds – indicates market urgency. Organisations are not waiting to see how the technology develops; they are buying now.
What to do this week
If you’re building automation that takes actions on your behalf – sending messages, updating systems, making decisions based on data – prepare internal documentation of risk points. Where can the system do something you didn’t intend? Where do you lack a stop point?
Business clients are starting to ask these questions even for smaller automations, especially if they touch customer data or financial processes. If you work with an external vendor, ask them how they monitor unexpected behaviour. If you’re building it yourself, add a logging layer that records the main steps in the process – not just the final result.
Companies already running AI agents for business should check whether they have the ability to stop an action mid-process, not just at the end.
Frequently asked
No, if you use models through their regular interface without automation. Platforms like Onyx are designed for organisations building agents that perform automated actions in internal or external systems, not for manual use of models.
Regular AI security focuses on protecting the model itself – for example, preventing prompt injection or data leakage. Agent security focuses on the actions the system takes after it makes a decision – for example, preventing file deletion or sending unauthorised messages.
A short call, no commitment, and we will tell you whether it is relevant to your business or not.