Journal
AI Agents Are Creating a New Security Perimeter Inside the Enterprise
Oct 19, 2026 · Donoghteh

As employees build and deploy AI tools at speed, security teams face a shift from monitoring user behavior to governing agent behavior. Pluto Security’s $42 million raise highlights the emerging opportunity and the operational questions buyers must answer.
The enterprise security boundary is moving
The spread of generative AI has changed how software enters the workplace. A team can adopt a new assistant, connect an external service, or build a lightweight internal application long before a formal procurement process catches up. The result is not simply more software. It is software that can interpret instructions, access data, and take actions on behalf of people.
That shift is drawing investor attention to a new category of enterprise security. Pluto Security emerged from stealth in October 2026 with $42 million in funding, according to TechCrunch: an $8.5 million seed round led by TLV Partners and a $34 million Series A led by Jump Capital. The company says it helps organizations discover and manage employees’ AI tools, agents, connectors, and related components.
The financing is a signal that buyers and investors see a governance gap. It is not, by itself, proof that the category has settled on a winning product or that every organization needs the same controls. The durable opportunity will depend on how well security products reduce real risk without making useful AI impractical to adopt.
From user behavior to agent behavior
Traditional security programs have largely been designed around people, devices, applications, and networks. AI agents complicate that picture. An agent may use a person’s credentials, call tools, access a connector, or chain multiple actions together. Understanding the initiating user is necessary, but it may not explain what the system actually did, what data it touched, or whether its actions stayed within the user’s intent.
TechCrunch reports that Pluto monitors AI tools and the components they rely on, including skills, extensions, MCP servers, and applications created by employees. The company says it can analyze activity and enforce policies, including intervening when an agent attempts to obtain permissions beyond those intended. These are company descriptions of its product; organizations evaluating such claims should validate them in their own environments.
This is a meaningful design challenge. A control that only blocks tools can push usage into less visible channels. A control that observes everything but cannot distinguish normal work from a dangerous action may flood security teams with noise. Effective governance needs context: who initiated the action, which data and permissions were involved, what the agent was trying to accomplish, and whether the action can be safely paused or reversed.
The visibility gap starts outside engineering
One detail in the reporting is particularly instructive: Pluto’s CEO said that about 60% of the risks the company detects across its customer base are associated with tools built by non-developers, with finance, marketing, and legal teams among the sources. That figure is attributed to the company and should be treated accordingly. Still, it illustrates a broader reality. AI-enabled development is spreading beyond teams that traditionally own software delivery.
A finance analyst may create a small application to manage customer information. A marketer may connect an AI service to campaign data. A lawyer may test a workflow assistant with sensitive documents. These experiments can be valuable, but they can also create exposed data, excessive permissions, or unclear ownership. Security cannot assume that risk only arrives through the engineering organization.
The answer is not to discourage experimentation. It is to make safe experimentation easier than unsafe improvisation. That means maintaining an inventory of tools and data flows, offering approved components, defining sensible permission boundaries, and giving employees clear routes to test new workflows.
What buyers should measure
For security leaders, the first question is not whether a vendor has an AI label. It is whether the product gives the organization actionable visibility and control. A focused evaluation should test several things:
- Coverage: Can the system discover relevant AI applications, agents, extensions, connectors, and locally built tools across the actual environment?
- Context: Can it distinguish a routine action from one involving sensitive data, unexpected access, or a risky tool chain?
- Intervention: Can it pause, block, or request confirmation for a clearly defined high-risk action without breaking ordinary work?
- Evidence: Does it provide an understandable record of what happened, who initiated it, which resources were involved, and what policy applied?
- Operational fit: Can security teams investigate alerts and tune policies without creating an unsustainable queue of false positives?
- Privacy and governance: What information is collected, where is it processed, how long is it retained, and who can inspect it?
These questions should be tested against representative workflows, not just a polished demonstration. Measure time to discover a risky component, the proportion of alerts that lead to useful action, false-positive rates, time to contain an incident, and the amount of employee work disrupted. Track whether teams can resolve issues through approved alternatives rather than simply losing visibility.
A new control layer, with familiar trade-offs
Pluto’s reported early traction, including more than 30 enterprise customers and over $2 million in annual recurring revenue, points to buyer interest. Those figures come from the company as reported by TechCrunch, and early revenue does not establish long-term retention or broad market fit. For the category as a whole, the harder tests will be renewal, expansion across departments, measurable risk reduction, and the ability to keep pace as AI products and agent frameworks change.
There are also important trade-offs. Monitoring prompts and activity can expose sensitive employee or customer information. Broad collection may help reconstruct an incident but can create its own privacy and access-control risks. Inline intervention can prevent a dangerous action, but poorly tuned controls can interrupt legitimate work. Buyers should assess these concerns explicitly and involve security, privacy, legal, IT, and the teams expected to use AI.
The strongest products will likely combine discovery, policy, and response with a transparent operating model. They will help organizations set boundaries before incidents occur, explain why an action was flagged, and provide a practical path to resolve the issue. Human review should remain available for uncertain or consequential cases.
What founders and investors should watch
For founders, the opening is not merely to build another dashboard. It is to become part of the workflow through which organizations approve, observe, and safely expand AI use. Products that can connect a risky action to a clear policy and a low-friction fix may deliver more value than broad inventories alone.
For investors, funding rounds are evidence of conviction, not evidence of durable adoption. Useful indicators include repeat usage by security teams, expansion from initial pilots into multiple business units, time saved during investigations, reduced exposure to high-risk actions, and the rate at which employees can move from unapproved tools to governed alternatives. It will also matter whether a vendor can maintain reliable coverage across a fast-changing ecosystem without collecting more data than customers are comfortable sharing.
The security perimeter is becoming less about where a person logs in and more about what a system can do once it is given context and tools. AI agents may increase productivity, but they also make permissions, traceability, and intervention more consequential. The companies that earn trust will be those that protect the ability to experiment while making the boundaries of action clear.
The next test is operational trust
The move from user behavior to agent behavior is more than a new slogan for cybersecurity. It reflects a real change in how work is performed and how risk can emerge. As organizations adopt agents, they will need controls that understand actions in context, preserve evidence, and intervene proportionately.
Pluto Security’s $42 million raise is one sign that this problem is becoming investable. The more important question is whether products in the category can reduce risk in measurable ways while preserving the speed and usefulness that made AI attractive in the first place. In enterprise AI, the security layer will be judged not by how much it can observe, but by whether it helps organizations act with confidence.