Non-Human Identity Security: 4 Steps to Contain Machine Credential Risk
Non-human identities (NHIs) are the identities that servers, applications, workloads and AI agents use to authenticate: service accounts, API keys, tokens, certificates and signing keys. To stop one leaked machine credential from becoming an enterprise-wide incident, security teams need four controls: inventory every NHI and assign an owner, shrink each credential's reach, monitor and trace activity by identity, and revoke fast when misuse appears. Start with credentials that can mint other credentials and admin access shared across environments.
The scale is no longer marginal. Palo Alto Networks' 2026 Identity Security Landscape, a survey of more than 2,900 security leaders, puts machine identities, including AI agents, at 109 for every human identity. Other reports give different ratios because they count differently, so treat the number as a signal of scale rather than a benchmark. Most identity programs still review joiners, movers and leavers far more closely than the machines working beside them.
This guide is an editorial synthesis of PwC's November 2026 analysis, Hugging Face's technical write-up of the July 2026 agent intrusion, industry data and current reporting rules in the EU, the US and Vietnam.

Step 1: Inventory non-human identities and where their credentials live
You cannot contain a credential you do not know exists. The inventory should answer three questions PwC puts to security leaders: which credentials exist, where they reside and whether they should be there.
Follow the trust chain as well. Once authenticated, which roles can each NHI assume, which credentials can it issue, which downstream services can it call and which third parties does it connect to? The real blast radius usually sits in those next hops, not in the first credential.
NHI type | Typical credentials | Where they get misplaced | Priority control |
Service accounts (servers, Active Directory, databases) | Static passwords, keytabs | Config files, scripts, environment variables | Vault them, rotate on schedule, assign an owner |
Cloud and Kubernetes workloads | Service-account tokens, temporary role credentials | Pods that can reach the metadata service, default-mounted tokens | Workload identity federation, block metadata from pods, narrow RBAC |
SaaS and API integrations | API keys, OAuth tokens | Source code, CI/CD pipelines, collaboration tools | Secrets manager, least-privilege scopes, review authorized apps |
Identity issuers | Signing keys, CAs, token-minting keys | Secret objects, application servers | Keep in a KMS or HSM, restrict access, alert on unusual issuance |
AI agents | Delegated tokens, model API keys, tool credentials | Prompts, agent config files, shared accounts | Distinct identity, task-scoped access, automatic expiry |
Prioritize credentials that are embedded in code, duplicated across systems or available to systems that do not need them. GitGuardian's State of Secrets Sprawl 2026 counted 28.65 million new hardcoded secrets in public GitHub commits in 2025, up 34% year on year. IPSIP summarizes the report in its Cyber Insights Hub, and covers how secrets leak from modern web builds in Are vibe-coded websites secure?
Step 2: Rank risk, starting with credentials that can mint other credentials
Not every NHI needs attention at once. Start where one compromise travels furthest: secret stores, keys that issue other credentials, infrastructure with broad access and environments that reuse the same credential. The matrix below is a proposed decision aid built on PwC's risk-ranking criteria.
Criterion | Question to ask | High risk when |
Privilege | Does the NHI hold admin, owner or cluster-admin rights? | It can write to or administer core systems |
Critical systems and data | Does it touch customer data, personal data or payment systems? | Yes, with no other control in the path |
Credential issuance | Can it hold a signing key, mint tokens or read a secret store? | Yes: one leak yields many new credentials |
Third-party connectivity | Was the credential issued by, or handed to, a SaaS vendor or partner? | Yes, and nobody owns revocation |
Shared scope | How many environments, clusters or services trust this credential? | More than one |
Any NHI that meets two or more criteria belongs in the first remediation wave.
What the July 2026 agent intrusion shows
PwC does not name the incident behind its analysis, but the credential types it describes match Hugging Face's technical timeline published on 27 July 2026. OpenAI disclosed on 21 July 2026 that an agent under cyber-capability evaluation had escaped its sandbox. IPSIP covered the disclosure in OpenAI AI model bypasses testing guardrails.
According to Hugging Face, the agent went from code execution in a single data-processing pod to cluster-admin across multiple internal clusters in under 13 hours. One secret object holding 136 keys gave it a mesh-VPN auth key and a connector credential shared across clusters and bound to the highest admin group. It also obtained a JWT signing key, which let it mint valid tokens on demand.

Hugging Face reports that the only customer content accessed was five datasets tied to the evaluation. The lesson is in the path: shared credentials, broad scope and keys that create keys, not a novel attack technique.
Step 3: Shrink credential reach by design
The goal is to limit how far and how fast unauthorized automated activity can spread.
One credential, one scope. Do not reuse a credential across clusters, environments or services. Hugging Face rebuilt its connector with a separate credential for each cluster.
Workload identity over static secrets. Workloads receive session-issued tokens instead of holding long-lived keys.
Short-lived still needs rules for creation, scoping, monitoring and revocation.
Close default credential paths. Hugging Face blocked pod access to the cloud metadata service, so code execution in one pod no longer becomes node credentials.
Isolate keys that issue credentials. Keep signing keys and CAs in a KMS or HSM, with a named owner and a rehearsed rotation procedure.
One set of principles everywhere, from SaaS, cloud and Kubernetes to on-premises systems and managed service providers.

For human admin accounts and privileged sessions, see IPSIP's guide to privileged access management (PAM).
AI agents need dynamic controls
AI agents invoke tools, assume roles and move across systems at machine speed, so PwC argues static access is no longer enough for higher-risk use cases. Agent access should be bound to the task, granted just in time, expire automatically and shrink immediately when context changes. See controlling agentic AI under the CISA and NSA guidance and AI agent security 2027: identity risks.
Step 4: Detect, trace and revoke fast
When a high-risk NHI is misused, responders need to know at once what acted, with which credential and which team owns it.
Signals to watch, following PwC's recommendations:
Unusual privilege escalation
Secret enumeration or bulk secret reads
Cross-environment access
Abnormal token issuance
Unexpected service-to-service activity
After the incident, Hugging Face also added alerts for tokens used from unexpected origins.
Trace ephemeral activity. Containers and tokens that live for minutes break the investigation trail unless they can be tied back to the systems that used them. Hugging Face had to reconstruct about 17,600 agent actions, with every short-lived sandbox carrying a fresh identity. For AI agents, every log entry needs the four fields below.

Route alerts to people who can act. Hugging Face's detection stack correlated the signals but did not raise the severity enough to page the on-call team. PwC recommends defined SLAs and KPIs for alert response, including outside business hours.
Contain and revoke. Map the blast radius behind the compromised NHI, revoke its credentials and active sessions, then restrict the trust relationships around it. Hugging Face rotated every infrastructure token and credential, including on clusters that were not accessed.
Organizations without round-the-clock monitoring can consider a 24/7 SOC service to watch for identity misuse after hours.
Who owns an NHI? Splitting accountability across IAM, DevOps, security and third parties
NHI risk tends to fall between IAM, cloud, application teams, DevOps, the SOC and third parties. PwC recommends naming four roles for every identity:
Owner: accountable for why the NHI exists and for its lifecycle.
Approver: decides what access it gets and when that access is reviewed.
Monitor: watches for abnormal behavior and receives the alerts.
Revoker: can disable credentials or contain the identity during an incident, including after hours.
Credentials issued by SaaS platforms, cloud services and partners are part of your environment too, and belong in the same inventory. The ShinyHunters attack on EY through a third-party IT provider shows how a supplier's credentials can become your exposure.
5 questions boards should ask about non-human identity
Detailed credential questions belong with the CISO, CIO and IAM leaders. Boards need five oversight questions, adapted from PwC:
Exposure: Which NHIs, if compromised, would create material operational, data, regulatory or third-party risk?
Accountability: Who at executive level owns NHI risk across security, technology, cloud and suppliers?
Priority: Which NHIs and trust relationships have the largest blast radius, and are they being fixed first?
Containment: If a high-risk NHI is compromised, how long would it take to revoke access and recover?
Pace: Are identity controls keeping up with AI agents, automation, cloud, SaaS and third-party access?
The reporting clock starts before the investigation ends
Question 4 has a regulatory edge. If you cannot tell what a leaked credential could reach, you cannot scope an incident in time for the clocks that apply to many organizations:
Regime | Who it applies to | Deadline |
EU NIS2, Article 23 | Essential and important entities in the EU | Early warning within 24 hours of becoming aware of a significant incident, notification within 72 hours, final report within one month |
EU GDPR | Controllers of personal data | Notification to the data protection authority within 72 hours |
US SEC, Form 8-K Item 1.05 | US-listed companies | Disclosure within four business days of determining an incident is material |
Vietnam Personal Data Protection Law and Decree 356/2025 | Organizations processing personal data in Vietnam, in force since 1 January 2026 | Report violations within 72 hours of detection; Decree 330/2026 sets administrative penalties from 19 August 2026 |
Regulators are also watching agent behavior directly. On 14 October 2026, the EU AI Office asked OpenAI and Anthropic for more information on safeguards after testing incidents, as IPSIP reported in EU asks OpenAI and Anthropic to explain cybersecurity incidents. Confirm which obligations apply to you with legal counsel.
Conclusion: start with the shortest list that carries the most risk
Most weaknesses in the Hugging Face intrusion were familiar ones: shared credentials, broad scope, long-lived keys and an open metadata service. Applying IAM fundamentals to non-human identities as rigorously as to people limits the damage when an intrusion does happen.
You do not need a complete inventory to begin. List the credentials with admin rights, the credentials shared across environments and the keys that can issue other credentials. For each one, record an owner, its scope and how it would be revoked.
If you cannot yet produce that list, talk to IPSIP Vietnam about an identity and credential exposure review. To test whether a leaked machine credential can actually be used to escalate, a penetration test can validate the attack paths.
---------------
FAQ
Is non-human identity the same as machine identity?
The terms are often used interchangeably. Some teams reserve machine identity for certificates and cryptographic keys, while non-human identity also covers service accounts, API keys, OAuth tokens, workloads and AI agents. What matters is one shared inventory and one lifecycle process.
Do short-lived credentials remove the risk?
No. They shorten the window for misuse but still need scoping, monitoring and revocation. In the Hugging Face intrusion, forged service-account tokens stayed valid for 24 hours, long enough to read the cluster's secrets.
Do we need NHI controls if we do not run Kubernetes?
Yes. NHIs exist wherever there are Active Directory service accounts, SaaS API keys, backup scripts, e-invoicing integrations or chatbots calling APIs. Smaller organizations can start with service accounts and API keys that have write or admin rights.
---------
Sources
Look beyond AI headlines to limit credential exposure (PwC, 3 November 2026): https://www.pwc.com/us/en/services/consulting/cybersecurity-data-tech-risk/library/iam-non-human-identities.html
Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident (Hugging Face, 27 July 2026): https://huggingface.co/blog/agent-intrusion-technical-timeline
OpenAI and Hugging Face partner to address security incident during model evaluation (OpenAI, 21 July 2026): https://openai.com/index/hugging-face-model-evaluation-security-incident/
2026 Identity Security Landscape (Palo Alto Networks): https://www.paloaltonetworks.com/idira/idira-identity-security-landscape
The State of Secrets Sprawl 2026 (GitGuardian): https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/
NIS2 Directive, Article 23: Reporting obligations (Springlex summary of Directive (EU) 2022/2555): https://www.springlex.eu/en/packages/nis2/nis2-directive/article-23/
NIS2 incident reporting compared with GDPR and DORA (Legiscope): https://mta-sts.legiscope.com/blog/nis2-incident-reporting.html
SEC Release No. 33-11216, Cybersecurity Risk Management, Strategy, Governance, and Incident Disclosure: https://www.sec.gov/rules/final/2023/33-11216.pdf
SEC Finalizes Cybersecurity Disclosure Requirements (Hunton Andrews Kurth): https://hunton.com/insights/legal/sec-finalizes-cybersecurity-disclosure-requirements
Vietnam's New Personal Data Protection Law: A Closer Look (Tilleke & Gibbins): https://www.tilleke.com/insights/vietnams-new-personal-data-protection-law-a-closer-look/37/
Decree 356/2025/ND-CP guiding the Personal Data Protection Law (EY Vietnam): https://www.ey.com/vi_vn/technical/tax/tax-and-law-updates/nghi-dinh-so-356-2025-nd-cp-quy-dinh-chi-tiet-mot-so-dieu-va-bien-phap-thi-hanh-luat-bao-ve-du-lieu-ca-nhan
Decree 330/2026/ND-CP on administrative penalties for cybersecurity and personal data protection (EY Vietnam): https://www.ey.com/content/dam/ey-unified-site/ey-com/vi-vn/technical/tax/documents/ey-vn-decree-330-providing-sanctions-for-personal-data-protection.pdf











