Cyber & Risk Innovation Institute

Look beyond AI headlines to limit credential exposure

hero image
  • November 03, 2026

What security leaders can learn from high-profile AI incidents as non-human identities multiply

Key takeaways:

  • As more work is delegated to machines and agents, organizations require stronger oversight of non-human identities.

  • Machine credentials that are stored, trusted, or scoped incorrectly can turn a narrow foothold into enterprise-wide exposure.

  • CIOs, CTOs, CISOs, and IAM leaders should know which non-human identities exist, where their credentials reside, how far access extends, and whether automated actions can be traced to an accountable human owner.

  • Governing non-human identities should be elevated to an enterprise security and resilience priority, not treated merely as a credential management task.

Recent headlines highlight autonomous agents acting beyond their boundaries. But they overlook a security gap organizations should address to limit the impact of this and other types of compromise: the governance of non-human identities and the credentials they use. As more activity shifts to automation, poorly managed or overly broad credentials can amplify the impact of a compromise, especially when machine access can span multiple systems or environments. This is more than a data access/exposure issue. The consequences of unauthorized non-human identities rooting around in enterprise systems extend to operational disruption, third-party risk, regulatory impact, and weakened resilience.

One high-profile incident involving agents demonstrates the risk of poorly governed credentials. After gaining initial access, an agent discovered and used credentials associated with other non-human identities, including service-account tokens, cloud credentials, static passwords, signing keys, and other forms of machine authentication. These credentials allow automated systems to act on behalf of the organization and, if they’re compromised, they can give an attacker persistent, high-impact access across critical systems and data.

Yet many identity and access management programs still focus primarily on people: joiners, leavers, movers, access requests, and periodic reviews for human users. That essential work no longer covers the full risk surface. Machines, workloads, applications, APIs, agents, and third parties can assume roles, delegate authority, and traverse trust relationships across the enterprise. Organizations should move to governing an expanding, interconnected identity environment and away from isolated credential controls.

How machine credentials can widen exposure

The population of non-human identities is growing quickly. Some estimates suggest they outnumber human users by a ratio of about 100:1, and it’s only going to grow. Yet these identities often lack the visibility, ownership, and access review applied to workforce identities, which can leave unmanaged access paths across the enterprise.

In the incident described above, an agent acting outside its authorized boundaries compromised a server handling routine file processing. The agent encountered multiple forms of authentication credentials, including a Kubernetes service-account token, temporary cloud credentials, a static database password, and a mesh-network authentication key. Most consequentially, it also encountered a signing key that could issue credentials.

The credentials were not tightly bound to the systems they originated from and, in some cases, were broadly trusted, allowing the attacker to contact the main cluster. Given the way permissions were scoped, the same credential was trusted across multiple clusters and granted administrator-level access. From there, the same credentials opened paths to additional clusters across multiple environments. The credential and trust architecture allowed a compromise in one workload to expand across multiple security boundaries.

Heterogeneous technology environments widen that exposure. Most enterprises now operate across multiple clouds, SaaS platforms, Kubernetes, on-premises systems, managed service providers, and third-party platforms—each with different identity, credential, authorization, logging, and revocation models.

Containing that exposure can be difficult for your organization without a clear view of the credentials that have access to your systems.

Many organizations still struggle to answer core hygiene questions. What credentials exist? Where do they reside? Should they be there?

Credentials that are scattered, duplicated, or embedded in places they don’t belong create easy paths for attackers. Credential visibility should also extend across the trust chain. Once authenticated, what additional roles, credentials, services, or downstream systems can a non-human identity reach?

In the example above, automated defenses helped detect misuse, but the response still depended on identity context: what acted, which credentials were involved, and whether that activity could be traced to an accountable system or team.

Better credential placement, narrower scoping, shorter-lived authentication, stronger workload identity controls, and tighter trust boundaries can materially reduce how far and how fast an unauthorized automated activity can spread across your systems. For AI agents in particular, the audit trail should include who initiated the task, which agent and runtime acted, which identity or credential was used, and which downstream systems or services were invoked.

Five lessons for leaders

  1. Credential hygiene shapes the impact of a compromise.
    In the incident described above, a single compromised server yielded four high-value credentials, including one that could issue access across an entire cluster. Organizations should verify that credentials are stored appropriately, limited to the required permissions, and separated across systems.
  2. Ephemeral activity changes both control and investigation.
    Short-lived software containers (workloads that start and shut down automatically), accounts, and deployment credentials can serve real business needs. In some cases, temporary credentials are safer than static ones, but they still require clear guardrails because they can fragment the audit trail. Traditional investigation models often assume activity can be tied to a stable identity or system. In this case, the attack path involved thousands of ephemeral events, each with different footprints, making it far harder to reconstruct the path of compromise. Organizations should confirm that short-lived containers, accounts, and credentials can be traced to the systems that used them and the teams responsible so that security teams can reconstruct the activity.
  3. Third-party credentials are part of your environment.
    Organizations no longer manage credentials created only within their own systems. Hosted platforms, SaaS solutions, cloud services, and partners introduce credentials that can affect enterprise exposure. Security teams need visibility into credentials connected to external platforms, as well as those created internally.
  4. Ownership and accountability cannot be fragmented.
    Non-human identity risk often spans IAM, cloud, application teams, DevOps, security engineering, the SOC, and third parties. Organizations should assign an owner for each identity and specify who approves its access, monitors it, and can revoke or contain it when something goes wrong.
  5.  Agentic AI requires more dynamic identity controls.
    AI agents can act, invoke tools, assume roles, delegate work, and move across systems at machine speed. Static access models are increasingly insufficient. Higher-risk use cases require task-bound access, continuous authorization, runtime monitoring, and the ability to rapidly reduce or revoke privileges as context changes.

How to reduce credential exposure

Gain visibility and prioritize risk

  • Inventory non-human identities and map their credentials. Identify non-human identities across your environment and document the keys, secrets, certificates, tokens, and static passwords. Pay particular attention to credentials that are embedded in code, duplicated across systems, broadly scoped, or available to systems that don’t need them. Extend the inventory across the trust chain, including roles that can be assumed, credentials that can be issued, delegated permissions, downstream services, and third-party connections.
  • Prioritize the riskiest locations. Start where a compromise will likely have the most reach: credential stores, secrets that can issue other credentials, infrastructure with broad access, and environments with repeated credential use or poor visibility. Risk-rank non-human identities based on privilege, access to critical systems or data, ability to issue or obtain additional credentials, third-party connectivity, and potential blast radius.
  • Maintain continuous discovery and effective escalation. Keep the inventory of non-human identities and credentials current and reassess risk as privileges and trust relationships change. Validate log aggregation, alerting, and notification procedures across critical applications and infrastructure platforms. Your on-call teams should have defined service-level agreements (SLAs) and key performance indicators (KPIs) for alert response, including outside normal business hours.

Reduce exposure by design

  • Reduce credential reach. Limit permissions to what each identity needs. Avoid broad trust, unbound credentials, and reuse of the same credential across multiple clusters or services. Apply least privilege where possible. Scope trust relationships narrowly to limit lateral movement if a credential is compromised.
  • Adopt safer patterns for automated systems. Use workload federated identity and issued tokens where appropriate to reduce reliance on static credentials. This approach registers systems with the cloud provider and uses issued tokens rather than open credentials. Temporary credentials aren’t inherently risky but they do require clear rules for creation, scoping, monitoring and revocation. For higher-risk agentic use cases, consider task-bound, just-in-time and context-aware access that expires automatically when the task or session ends.
  • Establish consistent controls across heterogeneous environments. Apply common identity-security principles across SaaS, cloud, Kubernetes, on-premises infrastructure, managed services, and third parties, even when the underlying identity and credential technologies differ. Your organization should be able to understand and enforce access consistently across those trust boundaries.

Detect, investigate, and contain misuse

  • Strengthen forensics for ephemeral activity. Confirm short-lived accounts, containers, and credentials can still be traced to accountable systems or teams. When an attack generates thousands of ephemeral events with different footprints, your security teams require enough context to reconstruct how automated activity moved through the environment.
  • Build identity-aware detection and rapid containment capabilities. Monitor for unusual privilege escalation, secret enumeration, cross-environment access, abnormal token issuance, unexpected service-to-service activity, and other signs of identity misuse. If a high-risk non-human identity is compromised, teams should be able to identify the downstream blast radius, revoke credentials and active sessions, and restrict related trust relationships quickly.
  • Test misuse scenarios early. Bring your defensive teams, offensive teams, and identity specialists together to understand how automated access could be misused and which controls would limit exposure.

Embed governance and secure practices

  • Give developers guardrails. Provide the tools, controls, and processes that make secure engineering the easier path. Ephemeral credentials can serve real business needs, but they need guardrails that limit reach, preserve traceability, and prevent insecure storage or reuse.
  • Connect automation patterns to controls. Identify which automation patterns create the most risk, which controls reduce impact, and where to invest first. This gives your organization’s leadership a more concrete view of what to assess, remediate, and monitor.
  • Establish a cross-functional operating model. Define how IAM, cloud, application teams, DevOps, security engineering, the SOC, AI governance, and third-party risk share responsibility for discovery, ownership, approval, monitoring, containment, and life-cycle management of non-human identities.

What security leaders should communicate to the board and senior leadership

CISOs, CIOs, and IAM leaders should help their boards and senior leaders understand where credential risk could expand exposure and what to prioritize. Boards and senior leaders don’t need every technical detail. The more detailed credential and control questions should sit with the CISO, CIO, and IAM leadership teams, supported by executive oversight.

The following practical questions can help focus board and senior leadership discussions:

  • Enterprise exposure: Do we understand where compromise of a non-human identity or machine credential could create material operational, data, regulatory, third-party, or resilience risk?
  • Accountability: Is executive ownership for non-human identity risk clear across security, technology, engineering, cloud, and third parties?
  • Risk prioritization: Do we know which non-human identities and trust relationships have the greatest potential blast radius, and are we addressing those first?
  • Containment and recovery: If a high-risk non-human identity or credential is compromised, can we quickly identify the affected systems and relationships, revoke access, and recover without major business disruption?
  • Pace of change: Are our identity-security capabilities keeping pace with the growth of AI agents, automation, cloud, SaaS, and third-party access?

Widespread damage from a compromise involving misaligned agents is not inevitable. Strong IAM fundamentals, applied to non-human identities as rigorously as to people, give CISOs concrete steps to strengthen their defenses and limit the impact when an intrusion does occur.

Contact us

Sowvik Chakrabarty

Sowvik Chakrabarty

Principal, Cyber, Data & Tech Risk, PwC US

Swaroop Dhulpet

Swaroop Dhulpet

Director, Cyber, Data & Tech Risk, PwC US

Follow us