AI agent governance moves from nice-to-have to survival requirement
AI agent governance is the set of tools, policies, and runtime controls that define what autonomous AI agents are allowed to access, decide, and execute inside enterprise systems, and that continuously monitor, block, or correct their actions to prevent security incidents, policy violations, and unintended business outcomes. Enterprises are now discovering that deploying AI agents without this governance is less like rolling out another SaaS app and more like letting thousands of junior employees loose with admin keys and no supervision. That realization is behind two headline funding rounds: Obsidian Security has closed an USD 85m (approx. RM391m) Series D to secure non-human identities and AI agents across third-party applications, and Zenity has raised USD 125m (approx. RM575m) in a Series C to secure AI agents acting inside corporate systems. Capital is voting loudly: control the agents, or they will control you.
Obsidian bets on runtime guardrails for rogue AI agents
Obsidian is building its business on a hard number: non-human identities already outnumber humans by 144 to 1 inside enterprise apps, and that gap is widening as agent-based automation spreads. The company’s USD 85m (approx. RM391m) Series D is not about abstract AI hype; it is about runtime governance that can stop agents from escalating privileges, overreaching on data, or breaking policy on platforms like Microsoft Copilot, Google Vertex, Amazon Bedrock, OpenAI, and others. In plain language, Obsidian wants to be the kill switch and control plane for agents embedded deep in third-party systems. It already counts more than 100 customers spending over USD 100k (approx. RM460k) a year and at least 14 spending more than USD 1m (approx. RM4.6m) annually. That traction matters because, in its own words, Obsidian intends to be “the platform that protects enterprises from agents going rogue in third-party applications”.
The platform’s direction is clear: extend agent access governance to more ecosystems, including Anthropic’s Claude Code and Cowork, while giving security teams an inventory of every agent, MCP server, and large language model connected to critical applications. This is not mere visibility reporting; it is about restricting dangerous permissions on production data and blocking unsanctioned tool use before damage occurs. The motivation is grounded in real-world failures: security teams have seen agents trigger 13-hour cloud outages, delete decades of irreplaceable personal files, and bypass built-in app guardrails to wipe production databases. Obsidian’s strategy is opinionated and, arguably, correct: if you cannot see every agent, control its reach in real time, and standardize policy across multiple AI ecosystems, you have no meaningful autonomous system control at all.
Zenity makes the agent—not the model—the security boundary
Zenity’s USD 125m (approx. RM575m) Series C is a direct bet that the real risk in enterprise AI is not the large language model, but what self-directed agents do once they are plugged into business systems. Its investors—Norwest, alongside SoftBank Vision Fund 2, Hitachi Ventures, and LG Technology Ventures—are backing a view that screening prompts and auditing outputs is no longer enough when agents can “take actions, make decisions, and touch every aspect of the enterprise”. Zenity claims a different stance: continuous, real-time monitoring of what agents are doing behind the scenes, with the ability to block or alter actions when they drift from their original purpose. That is rogue AI monitoring in practice, not mythology. If an agent’s behavior starts to endanger clients, data, or operational integrity, Zenity inserts itself in the control loop.
The company plans to use its new capital to expand across Asia-Pacific, the U.S., and Europe, but the more interesting choice is where it is focusing: heavily regulated sectors such as financial services, telecommunications, health care, pharmaceuticals, and energy, where agents are increasingly deployed not only internally, but in customer-facing roles like support and ticketing. That is where AI agent compliance stops being a checkbox exercise and becomes existential. One quotable statement captures the stakes: “As our use of AI agents continues to grow, maintaining security, governance, and operational control is essential”. Zenity is essentially arguing that if you do not treat agents as first-class security subjects—at least as important as human users—you will be caught flat-footed by their mistakes.
Why enterprise AI security is shifting to agents as the new perimeter
These funding rounds are symptoms of a larger shift: enterprises are no longer experimenting at the edge—they are wiring AI agents into daily operations. Businesses are rolling out agents built on models like Anthropic’s Claude, OpenAI’s ChatGPT, and Microsoft Copilot Studio, then wrestling with how to govern their behavior once embedded in the third-party applications that run the business. At the same time, enterprises in Asia-Pacific and beyond are rushing to deploy agents that take actions and touch every aspect of operations. That combination is why AI agent governance and AI agent compliance are now hot markets instead of niche obsessions. Traditional security approaches, focused on users and static permissions, cannot keep up with agents that self-orchestrate workflows, call tools, and switch models on the fly. Without dedicated autonomous system control, each new agent is a potential shadow IT department with no accountability.
One telling detail is Obsidian’s decision to track every large language model powering agents in an environment and alert security teams when models are switched or substituted so only sanctioned models remain in operation. That is governance at the model layer, but applied through the lens of agents and MCP servers. In parallel, Zenity’s real-time monitoring and intervention on agent actions show that governance must exist at runtime, not as an after-the-fact audit. When agents can connect unsanctioned to critical systems, escalate privileges, or access data they were never meant to see, the perimeter effectively dissolves. The new perimeter is the agent’s intent, its allowed tools, and its real-world behavior. Ignoring that reality is not conservative; it is irresponsible.
Financial services and cross-functional workflows will decide the winners
If you want to see where AI agent governance platforms will either prove their value or fail, look to financial services and cross-functional enterprise workflows. Zenity is clear that its heaviest focus is on regulated sectors such as financial services, telecommunications, health care, pharmaceuticals, and energy, where agents are moving into both internal and customer-facing roles. Those environments combine complex approval chains, strict compliance demands, and high-value data—all ideal testbeds for enterprise AI security. At the same time, Obsidian is pushing hard into the Fortune 500 and Global 2000, aiming to become the standard for governing what AI agents can reach and do inside enterprise systems. Cross-functional workflows—spanning CRM, ticketing, ERP, and cloud infrastructure—are exactly where unsanctioned agents and hidden MCP servers create the scariest failure modes.
The likely outcome is a consolidation around platforms that can give security teams a single control point across multiple AI ecosystems, with shared network intelligence so that a risk discovered at one customer becomes a safeguard for all. In that world, rogue AI monitoring is table stakes; enterprises will demand clear policy enforcement, auditable decision trails, and tight integration with existing compliance processes. The conclusion is blunt: AI agents will keep spreading through financial and cross-functional workflows whether or not security is ready. Enterprises that treat AI agent governance as an afterthought are betting their business on the hope that autonomous systems will behave. The smarter bet is to assume they will not—and to invest in the infrastructure that can keep them in line.
