The non-human identity problem is not new. The urgency around it is.
That was the through-line of a recent Clarity Security webinar bringing together Lalit Choda, founder and CEO of the Non-Human Identity Management Group and one of the industry’s most prominent voices on NHI security, and Alexis Moyse, CEO and co-founder of Clarity Security. The conversation covered AI agent governance, ownership accountability, observability, runtime enforcement, and where security teams should actually start — with the kind of directness that comes from two people who have been watching the same problems compound for years.
What follows are the key themes from that discussion.
The NHI Problem Is Older Than AI — Agents Just Made It Impossible to Ignore
Lalit Choda first start working with non-human identities over 25 years ago in an investment banking environment. The NHI challenge wasn’t new then either; what has changed is the scale.
“In the last 10, 15 years with cloud, SaaS-based services, and containers, the problems just got significantly worse,” Choda said. “We’re in a much more hyper-fragmented environment — multi-cloud, on-prem — and the way we’ve managed credentials that support non-human identities has typically been poorly managed and governed.”
The arrival of AI agents didn’t create the problem. It amplified it, and put it in front of everyone who had been comfortable ignoring it. “If you don’t get your foundations right, you’re going to be on a slippery slope as you start building out your agentic infrastructure,” Choda said.
Who Owns an AI Agent When Something Goes Wrong?
Ownership is one of the most debated questions in NHI governance, and the conversation surfaced a perspective that cuts through a lot of the noise.
“A lot of people focus on needing a human owner — maybe the person that built the agent,” Choda said. “I think that, as a core principle, is wrong. It should be whoever owns the product, the application, the workload, the agent itself. So it’s the team.” Mapping ownership to an individual creates continuity risk every time someone changes roles or leaves. Mapping it to the team that owns the underlying application preserves accountability through transitions.
The IAM team has a distinct but complementary role — not ownership in the business sense, but responsibility for making sure the full end-to-end lifecycle is managed correctly for those identities.
Moyse pushed the question further: what happens when an agent does something catastrophic? Who is accountable when an AI agent deletes a production database or causes a significant incident? The answer, both agreed, is that the distinction between ownership and accountability gets tested hardest when something breaks — and organizations that haven’t drawn that line clearly before an incident will be drawing it under pressure after one.
You Can’t Govern What You Can’t See — Why Observability Has to Come First
When asked to describe the North Star for NHI and AI agent security, Moyse was direct: a unified governance plane covering human identities, non-human identities, and agents — with everything visible, everything governed by a lifecycle, and real-time enforcement as the operating model.
Her answer on where to start: observability. “90% of IT and security typically starts with observability. How do you build solid observability across human, non-human, and agent identities and have that in a single pane of glass?”
Choda’s guidance aligned: understand what you have first. How many agents and NHIs are in your environment? What can they access? What APIs, databases, and downstream systems do they connect to? What credentials are they using, and who owns them? “Shadow AI is probably out of control in many organizations,” Choda noted. “If you can start to understand what you’ve got, then you can start managing the risk. Because you can’t stop it, but you can control it.”
The progression Moyse laid out — observe, triage, act — reflects how mature security programs build toward real-time enforcement without breaking the business in the process.
Runtime Enforcement: Why Prompt Instructions Aren’t Security Controls
One of the sharpest exchanges of the conversation came around the question of whether governing agent behavior is the same as governing agent access. The answer, clearly, is no — and the PocketOS incident anchored the discussion.
“You cannot rely on guidance that you give to your agent — your .md files, whatever,” Choda said. “You need to have clear enforcement at runtime.” The PocketOS scenario, in which an agent proceeded to delete a production database despite instructions not to perform destructive actions, illustrated the core problem: agents act to achieve goals, and they will take whatever path they determine is necessary to do so, regardless of the instructions they were given.
The implication is significant. For every action an agent decides to perform — connecting to another agent, accessing a database, calling a service it has never touched before — there needs to be a real-time assessment of whether that action is appropriate given the agent’s defined purpose and scope. If the action is appropriate, just-in-time ephemeral credentials with only the privileges needed for that specific action should be granted. If it isn’t, it should be blocked.
Moyse offered a reframe that’s useful for security teams trying to communicate the risk internally: treat AI agents the way you’d treat a potentially adversarial actor — not because agents are malicious, but because they’re unpredictable. “We have absolutely no idea what they might do, and they could go rogue.” That framing shifts the conversation from enabling agents to governing them, which is where security teams need to be.
Choda made the insider threat dimension explicit. “I actually think the bigger issue is going to be the internal threat from people using agents that try and please — want to do things and help you out — whether it’s your vibe coding agent or an agentic process where they try their best to do something they think is correct.” The agent equivalent of a well-meaning but unauthorized action taken by an overeager employee — except agents don’t pause to ask for clarification, and mistakes surface only after the damage is done.
Where Security Teams Should Start Today
The practical guidance from both speakers converged on a few clear priorities.
Start with discovery and know what you have. Inventory every NHI and AI agent in the environment, understand what each one can access, and identify who owns it. This is the foundation everything else builds on, and most organizations have less visibility here than they believe.
Build with the right foundations from the start. “Don’t make the mistakes we made with NHIs,” Choda said. “Dynamic, ephemeral, just-in-time — those things are mandatory. Don’t build your agents off legacy NHI processes, static credentials. Use best-of-breed practices as you build out your agents.” Organizations that hard-code credentials into agents or rely on long-lived static API keys are replicating the same governance failures that have defined the NHI problem for the last decade.
Alexis offered a hands-on exercise that requires no technical background: spin up an AI agent with full permissions and watch what it does. “At the end, ask it to print out the permissions it used and gave itself through the process. See how fast it grows its own permission set, everything it can do beyond the scope of the prompt as you wrote it.” The exercise is designed not to alarm but to build the kind of intuitive understanding of agent behavior that makes governance decisions more concrete.
Communicate risk in terms the board already understands. Choda’s advice on making the case upward: connect agents to the risks executives are already familiar with — data exfiltration, competitive exposure, production incidents. Moyse’s addition: break the ask into the components boards are already used to funding, rather than presenting a broad “govern AI” budget request. “Boiling the ocean is very difficult,” she said, “unless you just got breached.”
The Bottom Line
The NHI and AI agent security problem is not on the horizon. It’s already in most organizations’ environments — in the form of ungoverned service accounts, unrotated API keys, shadow AI deployments, and agents operating with more access than anyone intentionally granted them.
What Choda and Moyse made clear is that the path forward isn’t waiting for the category to mature or for a single platform to solve everything end to end. It’s starting with what’s knowable today — full discovery, ownership accountability, and observability — and building toward runtime enforcement and just-in-time access as the governance model matures.
The organizations that get their foundations right now will be equipped to scale their agentic programs securely. The ones that don’t will be dealing with the consequences faster than any manual process can track.
This post summarizes key themes from Clarity Security’s webinar “Beyond Human: NHI and AI Identity Governance,” featuring Lalit Choda, founder and CEO of the Non-Human Identity Management Group, and Alexis Moyse, CEO and co-founder of Clarity Security. Watch the full recording here.
