Israel's Bay emerged from stealth on November 10, 2026, with 50 million dollars in total funding and a system that checks AI agents' actions before they execute. The Tel Aviv company targets the point where a software assistant stops proposing an answer and starts changing something: opening access, running a tool or updating information in a business system.
According to Bay's announcement, the funding comprises two rounds: a previously undisclosed 13 million dollar seed co-led by YL Ventures and Striker Venture Partners, and a 37 million dollar Series A led by Battery Ventures. Micron Ventures, Microsoft's M12 fund and Wiz CEO Assaf Rappaport also participated in the Series A. Bay did not disclose its valuation.
That distinction matters: Bay is announcing its total funding, rather than another 50 million dollars on top of the seed. Behind the financing is an operational question: how do you give an agent enough freedom to work without turning every permission into an open invitation to perform any available action?
Checking the action itself
Bay says its system assesses each proposed action against the agent's assigned task and its previous activity in the session. It can allow the action, block it or request human approval. The check therefore concerns both the user's initial instruction and the next step the agent wants to take inside the working environment.
This reflects a practical distinction between technical permission and relevance to a task. Access to a system may enable many actions, although only some are needed for a particular request. In principle, an action is not justified simply because a tool can execute it. Its purpose, affected information and place in the workflow also matter.
Human approval provides another boundary. Pausing for clarification need not mean the agent has failed; it can reflect a limit the organization deliberately imposed. The usefulness of that mechanism also depends on the quality of its decisions and the explanation a person receives, beyond the mere presence of an approval button.
Three components, from discovery to policy
SiliconANGLE describes three main components. Wave scans a machine to inventory agents, extensions and local models. Bay Enforcer receives proposed actions through a native hook in the tool and returns a decision. Lighthouse drafts a separate policy for each team using its previous 30 days of agent activity.
In that account, Enforcer is a static executable and the check runs locally, using the session's context. Bay says decisions arrive within milliseconds. That is a company performance claim, rather than an independent measurement; response time in a particular environment remains something a prospective customer needs to assess.
Testing a policy against the previous month's activity adds a useful step between writing rules and enforcing them. Security teams can see which actions would have stopped and decide whether those are the intended targets. Such testing may reveal overly broad rules, but cannot establish how the system will handle every future task.
Fast installation is only part of adoption
Bay says deployment takes less than 5 minutes through an organization's existing mobile device management or endpoint detection and response tools. Its proposition is to fit into infrastructure a customer already operates, reducing the need for a separate infrastructure project just to discover which agents are active.
Deployment time should be distinguished from the time needed to define working boundaries. Even when software installs quickly, an organization must decide who can approve exceptions, which systems need greater protection and how work resumes after a block. Those are organizational decisions that cannot be inferred from a vendor's installation figure.
When assessing AI agent security tools, a practical trial can start with a limited workflow and examples the team already understands. Test both an acceptable action and one the organization wants stopped, then check whether the explanation supports a useful response. This evaluates local suitability beyond a general demonstration or a speed claim.
Recognizable customers and vendor-reported results
Bay names Micron, Wix, ClickUp and Peloton among its customers and says dozens of organizations already use the platform. Micron has two distinct roles: customer and investor through Micron Ventures. That relationship adds commercial context to the financing, but does not make every product performance claim independently measured evidence.
Bay says Fortune 100 companies using its platform tripled their use of autonomous agents. Separately, SiliconANGLE reports the company's claim of a false-positive rate below 1%. Both describe results presented by the vendor, rather than a commitment that every customer will achieve the same adoption level or accuracy.
For a prospective buyer, the relevant questions include what counts as a false positive, which actions were measured and what happens when a dangerous action is missed. Increased adoption also needs a baseline: how many agents were running, what tasks they had and which approvals stayed with employees. A single figure cannot answer all of that.
An Israeli team tackling expanding permissions
Bay was founded in early 2026 by CEO Yuval Zacharia, Chief Product Officer Aner Morag and VP of R&D Boris Danilovich. According to the company, Zacharia led about 50 people in Israel's Unit 8200 before leading research and AI at Hunters. Morag was an early Noname Security employee; Danilovich joined Preempt Security early, before its acquisition by CrowdStrike.
Globes reports that Bay employs 35 people, most in Israel. The company says the funding will expand product and engineering teams alongside sales, marketing and customer teams in the United States and Europe. It is emerging publicly about 9 months after its founding, with financing to grow an operation already serving enterprise customers.
For security leaders, the significant proposition is connecting the purpose of a task with permission for the next action. That needs testing where it will actually operate: against an organization's tools, data and exceptions. Those trials will show how far agents' autonomy can expand and which decisions should remain with people.