When Hours Beat Quarters: Identity Governance Becomes the New Front Line of Vulnerability Management

For more than a decade, vulnerability management ran on a comfortable rhythm. Researchers disclosed a flaw, vendors shipped a patch, and security teams worked the service-level agreement: thirty days for criticals, sixty for highs, ninety for everything else. That cadence rested on a quiet assumption: the time between disclosure and weaponization was always longer than the time between patch and deployment. AI-assisted vulnerability discovery has erased that assumption. Recent peer-reviewed work demonstrates large language model (LLM) agents reproducing real-world vulnerabilities from their text descriptions, multi-agent systems exploiting zero-days without prior knowledge, and benchmarks that evaluate AI offensive capability across more than 1,500 real software flaws. The exploit-development side of the equation now moves in hours. The patch-and-deploy side still moves in weeks. Most enterprise programs are living inside that gap, and the only durable way out of it is a different theory of defense, one where containment, not patch velocity, is the deliverable.

The Containment Question Is an Identity Question

When a defender cannot reliably patch faster than an attacker can weaponize, the strategic question stops being how do we close every door? and becomes what does an attacker reach when one door fails? That is an identity question first and a network question second. Modern enterprise estates are flat in ways that matter: an over-privileged service account, a stale OAuth token, or a workload identity with a cross-account trust each becomes a horizontal expressway through what looks on a network diagram like a segmented environment.

The National Institute of Standards and Technology made the same shift explicit in February 2024 when it released Cybersecurity Framework 2.0, elevating GOVERN to a first-class function alongside Identify, Protect, Detect, Respond, and Recover. The framework's authors recognized what practitioners were already seeing: cybersecurity outcomes increasingly depend on decisions about authority, accountability, and identity that sit above any individual control. The CSF 2.0 governance layer is the place where AI-era access policy gets written down, and it is where most programs are still under-investing.

Non-Human Identities Are the Privileged Population Now

CyberArk's 2025 Identity Security Landscape, drawn from a survey of 2,600 security decision-makers, found that machine identities now outnumber human users by more than 80 to 1, and 42 percent of those machine identities hold privileged or sensitive access. Yet 88 percent of respondents define a privileged user as a human concept exclusively. The arithmetic is unforgiving: most of the privileged population in a typical enterprise is governed like an afterthought.

Agentic AI has accelerated the problem. An autonomous coding assistant with repository write access, an LLM-powered finance agent that posts entries to the general ledger, and a Model Context Protocol (MCP) research agent calling external tools each represent a new privileged identity class with capabilities that did not exist twenty-four months ago. Industry analysts project that a meaningful share of enterprise breaches by 2028 will trace back to AI agent abuse. Whether or not any specific forecast lands, the structural point holds. An identity that acts faster than a human, with broader scope than a service account and looser oversight than either, is exactly the access pattern attackers want to compromise.

The remediation is unglamorous and well understood, which is why programs keep deferring it. Every non-human identity needs an owner of record, a documented purpose, a credential lifecycle covering provisioning, rotation, and deprovisioning, and access reviews on the same cadence given to human accounts. Recent work in the IEEE Open Journal of the Computer Society on Zero Trust maturity argues a similar point from a different angle. Identity is the foundational control vector, and treating workload identity, automation tokens, and human users as one governable population is what makes Zero Trust an architecture rather than a slogan.

The Asset Graph Has to Speak Identity

The single most useful question during an incident is also the one most asset-management programs cannot answer in under an hour: which identities can reach this asset, and which assets can this identity reach? The reason it is hard is structural. The configuration management database typically lives in one tool, identity and access management data lives in another, and the join key between them is rarely consistent. A traditional CMDB built for change management was never designed to answer reachability questions at incident speed.

The fix is to make runtime telemetry the source of truth and to join asset and identity data at ingestion. Cloud asset APIs, endpoint telemetry, Kubernetes control planes, and the identity provider all already carry the relevant data; cyber asset attack surface management platforms or a custom graph database can ingest it. The bar a security engineering team should clear is whether it can run a reachability query in minutes during a live incident. If the honest answer is we would have to pull a few CSV exports and pivot in a spreadsheet, the program has a foundational gap that no amount of additional perimeter spend will close.

Detection Has to Move Left of the Endpoint

Endpoint detection and response remains essential, but it fires at exploitation or post-exploitation. When exploitation can complete in seconds, that is too late as a primary detection layer. Identity threat detection and response capabilities (alerts on anomalous OAuth grants, impossible travel, new federation trusts, dormant service-account reactivation, or unexpected privilege escalations) fire earlier in the kill chain, sometimes by hours. The investment thesis is straightforward: signals that arrive before code execution are worth more than signals that arrive after.

Behavioral baselines for autonomous agents deserve the same engineering attention that programs gave EDR a decade ago. For each agent with production write access, the program should establish baselines for repositories touched, endpoints called, and actions taken per hour, then alert on deviation. Recent academic work on prompt injection in AI security agents demonstrates that compromised agents can manipulate the very systems meant to oversee them. That makes treating agent telemetry as adversary-resistant, including write-once logging, separation between agent and audit identities, and out-of-band integrity checks, a near-term requirement rather than a forward-looking nice-to-have.

Governance Is What Lets Defenders Move at Machine Speed Safely

The hardest part of the AI-era reprioritization is rarely technical. It is the negotiation that lets a security organization act at machine speed without manufacturing self-inflicted outages. The shape of the answer is a written framework of pre-approved containment authorities: a list of actions (block an IP at the edge, isolate a single endpoint, revoke a single token, fail a single build, roll back a single deployment) where each entry carries a defined trigger, a bounded scope, a named approver, an audit cadence, and a documented rollback path. Actions that cannot be cleanly scoped and reversed do not go on the list; they stay in the human-in-the-loop queue where they belong. This artifact is what converts the abstract phrase automated response into something a chief information officer and general counsel can sign.

Where to Start on Monday

1. Inventory every non-human identity with production write access. Agents, service principals, service accounts, and OAuth clients each need a human owner of record and a credential lifecycle. This list is the new crown-jewel registry; the rest of the program builds on top of it.
2. Stand up a reachability query against a joined identity-and-asset graph. A small Neo4j prototype fed from endpoint telemetry, the identity provider, and cloud asset APIs delivers more incident-hour value than waiting six quarters for a perfect commercial CAASM rollout.
3. Replace standing privileged access with just-in-time provisioning wherever workflows tolerate it. The privileges an attacker most wants should not exist in the precise moments those privileges would be useful.

The Next Eighteen Months

A decade of secure-software guidance from the IEEE Computer Society has emphasized that security has to be built in, not bolted on. The same logic now applies to identity. Programs that treat every agent, service account, and workload as a first-class governed identity, with an owner, a documented purpose, a lifecycle, and a behavioral baseline, will be the ones whose containment holds when the next exploit cycle compresses again. As earlier IEEE Computer Society analysis on AI in cybersecurity anticipated, the defenders who will fare best are those who pair AI-assisted detection with disciplined access governance, rather than treating either as a substitute for the other.

The programs that keep treating non-human identity as a low-priority cleanup item will keep finding out, the hard way, that the perimeter they thought they had was always a directory of trust relationships. The work to fix that is not glamorous, but it is buildable with technology that exists today and a budget that can be reclaimed from controls that already do not earn their keep. Closing the velocity gap between the attacker's loop and the defender's loop is fundamentally a program-design problem. Identity is where the design decisions land first.

About the Author

Priyanka Neelakrishnan, B.E., M.S., M.B.A. is a Director Product Manager at Veza Technologies that is acquired by ServiceNow, a published cybersecurity author, and a speaker whose work focuses on Enterprise Data Security and Product innovation. Connect on LinkedIn.

Disclaimer: The authors are completely responsible for the content of this article. The opinions expressed are their own and do not represent IEEE’s position nor that of the Computer Society nor its Leadership.