Get Identity Governance Right, and the Autonomous SOC Earns Its Autonomy

The network perimeter that security programs spent two decades hardening has effectively dissolved, but it hasn't been replaced by a cleaner boundary. It's been replaced by a sprawl of identities, most of them not human. Service accounts, API keys, workload credentials, and now autonomous agents all authenticate continuously and carry permissions that would set off alarms if a person held them. Defense now turns on a narrower problem: who or what is authorized to act at the exact moment an action happens, and whether anyone can prove it afterward.

Shain Jacob has spent more than twenty years securing large, regulated environments, and now oversees enterprise technology and information security for one of the largest state health agencies in the country. As Interim Chief Information Security Officer at Texas Health and Human Services, his mandate covers security architecture, cloud modernization, and compliance across HIPAA, NIST, and state frameworks, for systems that hold the data of millions of Texans. His view of the identity problem starts with a correction to how most teams frame it.

"Identity is where the majority of tool sprawl happens, and the network perimeter has effectively disappeared," Jacob says. "Without clear ownership of governance decisions, more tools doesn't mean anything."

The perimeter multiplied instead of moving

Jacob's central concern isn't identity in the abstract. It's the volume and opacity of non-human identities now operating inside enterprise systems. Knowing who has access to what, and whether that access still makes sense, is the baseline. The more challenging layer sits underneath it, where machine identities authenticate against each other with no person in the loop. Recent survey work from the Cloud Security Alliance found that most organizations have no clear ownership or accountability for their AI and machine identities, and that a large majority don't trust their legacy IAM tooling to manage the risk at all.

Service accounts are where Jacob has watched this go wrong most often. Because a shared account gets handled by multiple people, accountability dissolves the moment something breaks. His team has moved to eliminate them and tie every privileged identity back to a named person. "We now ensure there are no service accounts in play. Every one is associated to a named account, and we know who authorized it, when, where, why," he says. "We make sure it's either just in time, or a time-based one-time password is enabled for that particular session." The approach lines up with the zero trust principles NIST laid out years ago, which pull the security decision away from network location and onto the identity and the moment of the request.

Tooling alone doesn't close the gap, and Jacob is direct about the cost of assuming it will. "Every new platform is another integration that we have to maintain, another vendor to risk-assess, another source of logs competing for a security team's limited attention," he says. He still runs multiple scanners and redundant endpoint tools on purpose, so that a miss by one gets caught by another. "We don't rely on just one tool, because if one fails, the other one picks it up." What keeps that redundancy from compounding the noise is governance sitting on top of it. Someone has to decide which tool is active and which is passive, and that decision has to have an owner.

When authority gets delegated

Delegation is where the trust chain starts to fray. When a request passes through a service account, then an API, then a database, and on through several more machine identities, the logs stay accurate but accountability loses its grip. Jacob's answer is architectural. His team is building toward an agentic SOC layered on a medallion data model, with bronze holding immutable raw capture, silver holding normalized and vendor-neutral schemas, and gold holding the enriched entity timelines the security team actually investigates against. The point is total visibility into what's happening across the infrastructure, so that any deviation has a traceable path back to a source.

On top of that data, behavioral analytics watch for identities acting outside their normal pattern, and a library of playbooks flags anything that looks malicious or out of the ordinary. When something trips, an agent raises it, and a human steps in to confirm whether it's benign. "A human gets into the loop and then makes sure that, 'hey, this is nothing wrong, he was accessing GitHub or he was doing this or that,'" Jacob says. That human validation is fixed, not optional. Detection engineering and behavioral baselining are exactly the kind of work that keeps pulling the SOC's structure toward outcomes rather than console-hopping.

The agentic SOC's real ceiling

The capability is genuinely impressive, and Jacob is careful not to undersell it. The agentic SOC in his environment can write back to disable a compromised account and reach into the endpoint platform to contain an affected host automatically. He's chosen not to enable that autonomous containment yet, keeping a person on the trigger. The reason is accuracy, and he has the failure cases to justify it.

Given a set of source and destination IP addresses and asked to separate genuine work activity from malicious traffic, the agentic system has returned wrong answers. "It will tell that this IP address is a good IP, but in all reality it is a bad IP. It only looks at one or two databases," Jacob says. The agent streamlines correlation well. It doesn't yet clear the bar for judgment. "Validation, we need to do it ourselves." That gap is why triage and containment can move to machine speed while irreversible action stays human-directed, and it's the same tension that keeps surfacing when teams try to improve detection coverage without inheriting a flood of false confidence. For a regulated agency, a confident wrong answer that auto-contains a legitimate host is its own kind of incident.

That's also the argument for governance running in parallel with identity governance rather than trailing it. The reason to invest in an agentic SOC, in Jacob's framing, is speed against a threat landscape where the majority of attacks he now sees are AI-driven. Matching that speed safely is a governance problem before it's a tooling one, which is the through-line in how readiness frameworks are being built for agentic operations elsewhere.

More automation, more people

The expectation that autonomous tooling thins out the SOC runs opposite to what Jacob is seeing. Headcount is going up, not down. "As more and more AI tools are coming into play, the workload has increased drastically. Especially with AI identity coming into play. We need a lot more AI engineers in security," he says. The skill mix shifts toward agent governance, detection engineering, and oversight of autonomous workflows, but the net demand climbs.

Part of that pressure comes from how development has decentralized. AI has made it easy for people well outside the security org, in HR, sales, marketing, or on a product team, to build and ship applications. "It's very easy for people to build applications now, and how good is it? We cannot predict that," Jacob says. Stale or outdated records that never should have left a controlled system can end up pasted into public generative AI tools by someone who doesn't know better. Governing that requires an escalation model rather than a gate, and it puts the insider exposure problem in front of teams that were built to watch the perimeter.

Jacob does credit AI for real wins on the build side. "Some of the applications that no vendor was willing to touch, we were able to convert into a production application and modernize it," he says. He pairs every one of those wins with a hard caveat. "We just have to be very careful that there are no bugs and no backdoor open, so vulnerability scans are being done in a timely fashion. Cloud security is done, DSPM is done." That means the whole thing runs through a DevSecOps lifecycle meant to catch what speed introduces.

Foundations before autonomy

For all the attention on the agentic layer, Jacob keeps returning to the plumbing beneath it. Metadata tagging lets his team correlate where an incident happened and how the attack vector was exposed. Data lineage tells them where sensitive data originates and where it flows, so exposure can be traced and scoped when something goes wrong. The visibility problem he's most focused on is that most enterprises can see activity around the perimeter but have almost none into application and database logs. "One of the things that we are trying to do is get these basic foundational things set up so that these datas are being protected," he says. The agentic SOC and the volume of alerts it's meant to tame both rest on that groundwork. Get the lineage, tagging, and log coverage right, and the autonomy has something trustworthy to act on. Skip it, and every downstream decision inherits the gap.