Upwind acquired Aegis, a nine-month-old Israeli startup that developed a platform to identify attack patterns involving AI agents. Market sources estimate the deal at around $30 million in stock. Aegis had no customers or revenue at the time of acquisition.
What happened
Founders Omri Limor and Saar Anconina will lead Upwind AI Security Labs. Upwind, a cloud security company, identified a market gap: traditional security tools are not built to detect attacks that pass through AI agents. The acquisition happened just nine months after Aegis was founded, before the company had time to acquire customers.
The deal was in stock, not cash. This means Upwind acquired technology and a team, not a revenue-generating business. The speed of the transaction indicates the urgency of the issue from Upwind’s perspective.
Why it matters beyond the headline
A new security threat does not become relevant only when there is a major breach in the news. It becomes relevant when security companies invest resources in solving it. The fact that a cloud security company is acquiring technology specifically to detect attacks through AI agents means the threat is already in the field.
An Israeli organisation building an AI system needs to understand that threats are evolving faster than the technology itself. An AI agent can perform actions that appear completely legitimate—retrieving documents, sending API requests, accessing databases—but in practice leak information or perform unauthorised actions. Existing security tools look for malware signatures or anomalous traffic patterns, not anomalous agent behaviour.
This is not just an issue for banks or insurance companies. Any business building an AI system that interfaces with internal systems or sensitive data stores needs to ask these questions at the design stage.
What to do this week
If you are in the planning stage of an AI system, add one question to your requirements list: how will we detect if the agent is doing something it should not? This is not just a technical question, it is a question of defining boundaries. What is the agent authorised to do, and what is it not authorised to do, even if it is technically capable?
If you already have an AI system in production, check whether you have logging of every action the agent performs. Not just technical logs, but logging that allows you to answer the question: was this action legitimate in the business context? If you do not have an answer, this is a gap that needs closing.
For organisations working with security vendors, ask whether their tools detect anomalous AI agent behaviour. If the answer is not clear, this is a topic for discussion. AI agents for business require a security layer that matches how they work, not just the infrastructure they run on.
Sources
Frequently asked questions
Do regular security tools detect attacks through AI agents?
Traditional security tools look for malware signatures or anomalous traffic patterns, not anomalous agent behaviour. An AI agent can perform actions that appear completely legitimate—retrieving documents, sending API requests—but in practice leak information. This requires tools built specifically to detect anomalous agent behaviour.
When should I start thinking about AI agent security?
At the design stage, not after the fact. You need to define from the start what the agent is authorised to do and what it is not, and how you will detect if it exceeds those boundaries. The fact that a security company acquired technology to detect attacks through agents after just nine months indicates the threat is already here.