Non-Human Identity Statistics 2026-2027
By Axis Intelligence Research
Co-author: Marcus Chen | Last updated: November 7, 2026 | License: CC BY 4.0
Axis Intelligence Research’s Non-Human Identity Lifecycle Gap Index (NILGI) stands at 62.8 out of 100 in its November 2026 baseline reading. The gap is widest at both ends of the lifecycle: 82% of organizations found AI agents they did not know about in the past year, and 79% have no formal process to decommission agents.
Quick Answer
Non-human identities (NHIs) are the service accounts, API keys, OAuth tokens, certificates and AI agents that authenticate without a person. Palo Alto Networks counts 109 machine identities per human in 2026. Axis Intelligence Research estimates that a leaked, valid secret has a half-life of about 7.7 years, and its NILGI baseline reading is 62.8/100.
Key Findings
- Axis Intelligence Research finds that the Non-Human Identity Lifecycle Gap Index (NILGI) stands at 62.8/100 in its November 2026 baseline reading. It is built from five lifecycle-stage failure rates in two Cloud Security Alliance surveys.
- Axis Intelligence Research estimates that leaked machine credentials have a half-life of about 7.7 years. The estimate uses GitGuardian’s retests showing that secrets confirmed valid in 2022 fell from 70% still valid in January 2025 to 64% in January 2026.
- Axis Intelligence Research calculates that at least 73.5% of organizations claiming strong visibility into their AI agents also discovered unknown agents in the past year (Cloud Security Alliance, April 2026, n=418).
- Axis Intelligence Research projects about 123.7 machine identities per human by spring 2027 if the growth that Palo Alto Networks’ 2,930 respondents expect actually happens, up from 109:1 in 2026.
- Axis Intelligence Research calculates a 54-point gap between confidence and control: 91% of organizations believe they can rapidly contain a compromised AI agent, but only 37% apply credential revocation to autonomous agents (Palo Alto Networks, 2026).
How Many Non-Human Identities Are There per Human?
The count depends on who is counting and how. Survey-based and scan-based figures differ by design, so one number on a board slide hides a choice of method.
| Measure | Value | Basis | As of | Source |
|---|---|---|---|---|
| Machine identities per human identity | 109:1 | Survey of 2,930 security leaders, 20 countries | Mar–Apr 2026 | Palo Alto Networks, 2026 Identity Security Landscape |
| AI agent identities per human | 79 | Same survey | Mar–Apr 2026 | Palo Alto Networks |
| NHIs per human, environments scanned by Entro | 144:1 | Platform telemetry | H1 2025 | Entro Security Labs |
| Same measure, one year earlier | 92:1 | Platform telemetry | H1 2024 | Entro Security Labs |
Palo Alto Networks reports in its 2026 Identity Security Landscape that organizations manage 109 machine identities for every human identity. AI agents account for 79 of those 109, which is 72.5% (Axis Intelligence Research calculation, 79 ÷ 109). Ratio history, the composite of several published ratios and certificate-lifecycle data are covered on our machine identity statistics page, which owns these headline ratios.
Entro Security Labs measured 144 NHIs per human in the environments it scanned in H1 2025, up from 92:1 in H1 2024, according to its NHI & Secrets Risk Report. That is a 56.5% increase in one year (Axis calculation, 144 ÷ 92 − 1). Entro has been part of SailPoint since mid-2026, per SailPoint’s announcement dated June 29, 2026.
Marcus Chen: Don’t average 109 and 144. One figure is what CISOs believe they run. The other is what a scanner found in a population of environments that chose to install it. In both, the human is no longer the typical user of your identity provider. The NHI is.
What Will the NHI-to-Human Ratio Look Like in 2027?
Palo Alto Networks’ respondents expect AI agent identities to grow 85%, machine identities 77% and human identities 56% over the next 12 months. Applied to the 2026 baseline, Axis Intelligence Research projects about 123.7 machine identities per human by spring 2027 (109 × 1.77 ÷ 1.56) and about 93.7 AI agent identities per human (79 × 1.85 ÷ 1.56). These projections rest on respondents’ expectations, not on observed counts. They show how quickly an identity program has to scale just to stand still.
What Is the Non-Human Identity Lifecycle Gap Index (NILGI)?
The Non-Human Identity Lifecycle Gap Index (NILGI) is an Axis Intelligence Research metric that measures how often organizations fail at each stage of an NHI’s life, from creation to retirement. A reading of 0 would mean no surveyed organization fails any stage. A reading of 100 would mean every one fails every stage.
NILGI baseline reading: 62.8 / 100 (as of November 7, 2026)
| Lifecycle stage | Failure measured | Share of organizations | Source (survey, n, fieldwork) |
|---|---|---|---|
| Discover | Found AI agents they did not know about in the past year | 82% | CSA / Token Security, n=418, Jan 2026 |
| Own | No clear ownership or accountability for NHIs | 51% | CSA / Oasis Security, n=383, Aug–Sep 2025 |
| Govern | No documented, formally adopted policy for creating or removing AI identities | 78% | CSA / Oasis Security, n=383 |
| Respond | Over 24 hours to rotate or revoke a credential after potential exposure | 24% | CSA / Oasis Security, n=383 |
| Retire | No formal decommissioning process for AI agents (100% − 21%) | 79% | CSA / Token Security, n=418 |
| NILGI | Equal-weighted mean of the five stages | 62.8 | Axis Intelligence Research |
Formula: NILGI = (82 + 51 + 78 + 24 + 79) ÷ 5 = 62.8.
The inputs come from the Cloud Security Alliance’s State of Non-Human Identity and AI Security (released January 2026) and its Autonomous but Not Controlled agent survey (released April 2026). Both are online surveys of IT and security professionals run by the same organization, with samples of similar size. That is why their rates can sit in one index. Each was commissioned by an NHI vendor (Oasis Security and Token Security respectively). Vendors sell into the gaps they measure, but these remain the largest published samples that ask lifecycle questions directly.
Marcus Chen: The shape matters more than the 62.8. Response is the strongest stage: only 24% need more than a day to rotate a credential they know was exposed. The failures cluster where nobody is looking. Agents nobody registered (82%). Agents nobody shuts down (79%). Incident response works on the credentials you know about. The breach comes from the ones you don’t.
Does Other Data Confirm the Lifecycle Gap?
Yes. Platform telemetry and two independent surveys point the same way.
- Retire: 91% of former employees’ tokens were still active in Entro Security Labs’ 2025 State of Non-Human Identities and Secrets in Cybersecurity.
- Rotate: 59% of AWS IAM users, 55% of Google Cloud service accounts and 40% of Microsoft Entra ID applications had credentials older than one year, per Datadog’s State of Cloud Security 2025.
- Revoke: 64% of secrets GitGuardian confirmed valid in 2022 still worked in January 2026, per The State of Secrets Sprawl 2026.
- Rotate, survey view: 92% of organizations fail to rotate the majority of their NHI credentials on a 90-day cycle, and only 8% rotate more than 75% of them that often, per the 2026 SANS State of Identity Threats & Defenses Survey (March 10, 2026).
- Govern, breach view: only 46% of breached organizations secure NHIs in their AI workflows, per IBM’s Cost of a Data Breach Report 2026.
How Long Does a Leaked Secret Stay Valid?
About 7.7 years to lose half its validity, by Axis Intelligence Research’s estimate. GitGuardian retests the same cohort every year: secrets found on public GitHub and confirmed valid in 2022.
| Retest | Share of 2022 cohort still valid | Source |
|---|---|---|
| January 2025 | 70% | GitGuardian, State of Secrets Sprawl 2025 |
| January 2026 | 64% | GitGuardian, State of Secrets Sprawl 2026 |
| January 2027 (projection) | 58.5% | Axis Intelligence Research estimate |
The method is simple. Between the two retests, 64 ÷ 70 = 91.4% of the remaining valid secrets survived the year, so 8.6% were revoked or rotated. At a constant 8.6% annual revocation rate, half of a freshly leaked, valid secret population is still live after ln(0.5) ÷ ln(64/70) = 7.7 years. Carried forward, the 2022 cohort would still be 58.5% valid in January 2027 and would drop below 50% only about 2.75 years after January 2026, around October 2028.
GitGuardian’s own explanation is operational: teams cannot rotate secrets without risking production outages, “so they don’t.” Removing the leaked string from the repository does nothing to the credential. Revocation does.
Marcus Chen: Patch cycles are measured in weeks. A leaked API key’s useful life to an attacker is measured in years. That’s the number to put in front of an engineering VP who says the secret was “already deleted from the commit.”
How Fast Are New Secrets Leaking?
GitGuardian counted 28,649,024 new hardcoded secrets on public GitHub in 2025, up 34% year over year. AI-service secrets rose 81% to 1,275,105. These volumes, and their sector split, are tracked on our machine identity statistics page. Three figures from the same report matter most for NHI owners:
- Internal code is worse: 32.2% of internal repositories contained at least one secret, against 5.6% of public ones.
- Agent tooling leaks too: GitGuardian found 24,008 unique secrets in MCP (Model Context Protocol, the standard that connects AI agents to tools) configuration files, 2,117 of them valid.
- AI-assisted commits leak more: commits co-authored with Claude Code showed a 3.2% secret-leak rate, against a 1.5% baseline across all public GitHub commits.
One reconciliation for analysts tracking the series: GitGuardian’s 2025 edition counted 23.8 million new secrets for 2024. Its 2026 edition’s 34% growth rate implies a 2024 base of about 21.4 million (28,649,024 ÷ 1.34). The gap suggests that the 2024 count was restated. Year-over-year comparisons should therefore use each edition’s own growth rate rather than mixing totals from different editions.
How Many Organizations Have Unknown AI Agents?
82% discovered AI agents they did not know about in the past year, and 41% did so more than once, according to the Cloud Security Alliance’s April 2026 survey of 418 IT and security professionals, commissioned by Token Security.
| Agent governance measure | Share | Source |
|---|---|---|
| Discovered unknown AI agents in past 12 months | 82% | CSA / Token Security, Jan 2026 fieldwork |
| …more than once | 41% | Same |
| Believe they have strong visibility into agents | 68% | Same |
| AI agent-related incident in past 12 months | 65% | Same |
| Formal agent decommissioning process | 21% | Same |
| Automatically block out-of-scope agent actions | 11% | Same |
| Run fully autonomous agents | 13% | Same |
Among organizations reporting agent incidents, 61% cited data exposure, 43% operational disruption and 35% financial losses. Shadow agents most often appear in internal automation and scripting environments. Incidents are logged in our AI agent security incident tracker, and agent adoption is covered in AI agents statistics.
The Visibility Confidence Gap
The same 418 respondents answered both questions. If 68% claim strong visibility and 82% found unknown agents, at least 50% of the sample must fall in both groups (68 + 82 − 100). Axis Intelligence Research therefore calculates that at least 73.5% of organizations claiming strong agent visibility still found agents they did not know about (50 ÷ 68). This is a floor. The true overlap can only be higher.
Marcus Chen: That’s the finding I’d show an audit committee. It isn’t that visibility is low. It’s that the people who report high visibility are, in large part, the same people finding agents they missed. Confidence is measuring the dashboard, not the estate.
Can Organizations Revoke a Compromised AI Agent?
Mostly not as a routine control. Across AI agent types, 36% to 37% of organizations apply credential revocation, according to Palo Alto Networks’ 2026 Identity Security Landscape (n=2,907 for this question).
| AI agent type | Behavioral monitoring applied | Credential revocation applied | Source |
|---|---|---|---|
| Autonomous agents | 45% | 37% | Palo Alto Networks, 2026 |
| Conversational agents | 43% | 36% | Palo Alto Networks, 2026 |
| Generative agents | 40% | 36% | Palo Alto Networks, 2026 |
The same survey found that 99% of organizations already use AI agents and that 91% believe they can rapidly contain a compromised AI agent. Axis Intelligence Research calculates a 54-point gap between that belief and the 37% of organizations that apply credential revocation to autonomous agents (91 − 37). Revocation is the containment control: an agent whose credential cannot be pulled keeps acting at machine speed.
Breached organizations show the same pattern. IBM’s Cost of a Data Breach Report 2026 found that 46% of breached organizations secure NHIs in their AI workflows. Among those that do, 55% use machine identity inventory and lifecycle management, 39% secrets management systems, 36% behavioral monitoring, 32% zero trust extended to NHIs and 30% role-based access controls on service accounts. Axis Intelligence Research calculates that only about 25.3% of breached organizations run machine identity inventory and lifecycle management for NHIs in AI workflows (46% × 55%).
The SANS 2026 survey adds the governance view. 73% of organizations are deploying agentic AI or automations that require credentials, yet no single governance control, such as approvals, audit trails or sandboxing, is used by more than 40% of them.
Marcus Chen: Two surveys, two vendors, same shape. The CSA respondents think they see their agents and keep finding new ones. The Palo Alto respondents think they can contain their agents and mostly have no revocation path. The confidence number is the one that goes in the board deck. The control number is the one the attacker meets.
Are Non-Human Identities Over-Privileged?
It depends on the definition, and the published range runs from 19.4% to 97%.
| Measure | Value | Method | Source |
|---|---|---|---|
| NHIs with excessive privileges | 97% | Permissions granted vs. needed, Entro-scanned NHIs | Entro Security Labs, 2025 |
| AWS machine identities with full admin | 1 in 20 | Entro telemetry, H1 2025 | Entro Security Labs |
| AWS EC2 workloads overprivileged | 19.4% | Instance roles with risky data, lateral-movement or escalation permissions | Datadog, Sept 2025 telemetry |
| AWS third-party integration roles “dangerously overprivileged” | 12.2% (10% in 2024) | Vendor role can read all data or take over the account | Datadog |
| Machine identities with access to organizational data | 40% | Survey | Palo Alto Networks, 2026 |
| Orgs citing over-permissioned access as top NHI pain point | 51% | Survey | CSA / Oasis Security, Jan 2026 |
| Orgs that cannot enforce continuous least privilege for service accounts | 56% | Survey | Palo Alto Networks, 2026 |
| Orgs whose definition of “privileged identity” covers machine identities with sensitive-data access | 15% | Survey | Palo Alto Networks, 2026 |
| Privileged access managed through just-in-time or zero standing privilege | 39% (average) | Survey | Palo Alto Networks, 2026 |
The 97% and the 19.4% do not contradict each other. Entro counts any NHI holding permissions it does not use. Datadog flags only EC2 roles whose permissions would let an attacker read sensitive data, move laterally or escalate. The first measures hygiene. The second measures blast radius. For a risk register, the 1-in-20 full-admin figure and Datadog’s 12.2% of over-scoped vendor roles are the operational numbers. Each one names a specific identity that, if stolen, turns a tenant compromise into an account takeover.
The definition problem starts inside security teams. Palo Alto Networks found that only 15% of organizations define a privileged identity to include machine identities with access to sensitive data, while respondents say 40% of their machine identities already reach organizational data. An identity that is not classed as privileged does not get privileged-access controls. That is how over-privilege stays invisible to the privileged access management (PAM) program.
Third-party access is growing in parallel. AWS organizations deploy an average of 13 third-party integration roles, and 2.25% of those roles do not enforce an External ID, the control that stops another customer of the same vendor from assuming the role.
How Old Are Machine Credentials in the Cloud?
59% of AWS IAM users had an active access key older than one year, unchanged from 2024, in Datadog’s State of Cloud Security 2025, published October 8, 2025.
| Cloud credential | Older than 1 year, 2024 | Older than 1 year, 2025 | Change |
|---|---|---|---|
| AWS IAM user access keys | 59% | 59% | 0 pts |
| Google Cloud service account keys | 62% | 55% | −7 pts |
| Microsoft Entra ID app credentials | 46% | 40% | −6 pts |
Source: Datadog, State of Cloud Security 2025. Change in points is an Axis Intelligence Research calculation.
AWS is the only one of the three providers where the share of long-lived keys did not fall. Over half of those year-old AWS keys had also gone unused for more than 90 days. A key nobody uses but nobody revokes has all the risk of a live credential and none of the business value. On the workload side, 49% of EC2 instances enforced IMDSv2, the instance metadata service version that blocks the most common credential-theft path from a compromised workload. That share was 7% in 2022.
OWASP Non-Human Identities Top 10, by the Numbers
The OWASP Non-Human Identities Top 10 (2025) ranks NHI risks by exploitability, prevalence, detectability and impact. OWASP does not publish prevalence numbers for each risk. The table below is Axis Intelligence Research’s evidence map: the strongest primary measurement we found for each risk.
| OWASP rank | Risk (OWASP description) | Best primary measurement | Source |
|---|---|---|---|
| 1 | Improper offboarding | 91% of former employees’ tokens still active; 21% of orgs have formal agent decommissioning | Entro 2025; CSA Apr 2026 |
| 2 | Secret leakage | 64% of secrets valid in 2022 still valid Jan 2026; half-life ≈ 7.7 years (Axis estimate) | GitGuardian 2026 |
| 3 | Third-party NHIs | 92% of orgs expose NHIs to third parties; 12.2% of AWS vendor roles dangerously overprivileged | Entro 2025; Datadog 2025 |
| 4 | Insecure authentication | 2.25% of AWS third-party roles lack External ID enforcement | Datadog 2025 |
| 5 | Over-privileged NHIs | 97% hold excess privileges; 1 in 20 AWS machine identities is full admin | Entro 2025; Entro H1 2025 |
| 6 | CI/CD and cloud deployment authentication | 49% of EC2 instances enforce IMDSv2 | Datadog 2025 |
| 7 | Long-lived secrets | 59% of AWS IAM users hold a key older than one year; 71% of NHIs not rotated within recommended timeframes | Datadog 2025; Entro 2025 |
| 8 | Environment isolation | No primary prevalence measurement published | — |
| 9 | NHI reuse across applications | 60% of NHIs shared by more than one application | Entro 2025 |
| 10 | Human use of NHIs | No primary prevalence measurement published | — |
Two of the ten risks have no published prevalence data at all. For a category this heavily marketed, that gap is a finding in itself. Isolation and human misuse of service accounts are the two risks that need behavioral logs, not configuration scans, and nobody has published them at scale yet.
How Often Do Non-Human Identities Lead to Breaches?
90% of organizations suffered at least one successful identity-related breach in the past 12 months, and 76% suffered three or more, according to Palo Alto Networks’ 2026 Identity Security Landscape. The breach-level data tells a more complicated story.
| Breach measure | Value | Source |
|---|---|---|
| Breaches starting with vulnerability exploitation | 31% | Verizon 2026 DBIR |
| Breaches starting with credential abuse | 13% | Verizon 2026 DBIR |
| Breaches with third-party involvement | 48% | Verizon 2026 DBIR |
| Global average cost of a data breach | $4.99 million | IBM Cost of a Data Breach 2026 |
| Average cost of an AI-enabled malicious breach | $6 million | IBM 2026 |
| Share of malicious breaches that were AI-enabled | 1 in 4 | IBM 2026 |
| Orgs with an AI agent-related incident, past year | 65% | CSA / Token Security, 2026 |
| Average time to identify and contain a breach | 247 days | IBM 2026 |
| Average cost of a breach starting with abuse of valid accounts | $5.07 million | IBM 2026 |
In Verizon’s 2026 Data Breach Investigations Report, covering incidents from November 2024 to October 2025, exploitation overtook credential abuse as the top initial access vector, 31% to 13%. Exploitation now leads by a factor of 2.4 (Axis calculation). That does not acquit NHIs. The DBIR codes the first step of a breach. A stolen OAuth token or cloud key usually shows up later, in lateral movement and data access, after an exploit or a phished human opened the door. It also matters that third-party involvement reached 48% of breaches, because vendor integrations run on NHIs.
IBM’s 2026 Cost of a Data Breach study, based on 602 organizations breached between March 2025 and February 2026, puts the global average at $4.99 million. AI-enabled malicious breaches cost $6 million, a $1.01 million premium (Axis calculation). IBM also reports that more than 20% of organizations had a breach targeting AI models or applications, and that 92% of organizations with AI-related breaches lacked proper AI access controls. Breaches that began with abuse of valid accounts cost $5.07 million on average, and the full breach lifecycle averaged 247 days.
Which Identity Controls Lower Breach Costs?
IBM tests 30 cost factors, each in isolation against the $4.99 million average, so the savings below cannot be added together.
| Factor (IBM 2026) | Effect on average breach cost | Share of the $4.99M average (Axis calculation) |
|---|---|---|
| Identity and access management (IAM) | −$225,622 | −4.5% |
| Key lifecycle management tools | −$214,923 | −4.3% |
| Certificate lifecycle management tool | −$205,265 | −4.1% |
| Secret lifecycle management tools | −$163,668 | −3.3% |
| Supply chain breach (business partner compromise) | +$227,250 | +4.6% |
A supply chain breach adds about as much as an IAM program saves. For NHI owners, that pairing is the business case in one line: vendor integrations run on NHIs, and the controls that govern NHIs are among IBM’s top cost reducers.
When a breach starts, the response playbook often lacks NHI actions. Palo Alto Networks found that 41% of organizations include revoking tokens, API keys or certificates in their incident response procedures, and 41% include disabling user or service accounts.
Breaches are logged in our data breach tracker, and exploited vulnerabilities in the KEV tracker.
What Did the Salesloft Drift Breach Show About OAuth Tokens?
It showed that one integration’s tokens can open numerous customer tenants at once. Google Threat Intelligence Group reported that the actor it tracks as UNC6395 used compromised OAuth tokens from the Salesloft Drift integration to export data from Salesforce instances between at least August 8 and August 18, 2025. No Salesforce vulnerability was involved. The actor then searched the stolen records for AWS access keys, passwords and Snowflake tokens, so one stolen NHI was used to hunt for others. Salesloft and Salesforce revoked all Drift access and refresh tokens on August 20, 2025.
Marcus Chen: No zero-day here. An attacker held a vendor’s token, the token had more scope than the integration needed, and the stolen data contained more keys. That’s OWASP risks 3, 5 and 2 in one incident. The fix was the least glamorous control on the list: revoke everything and rotate what was exposed.
What Changes for Non-Human Identity Security in 2027?
Four dated signals frame next year’s numbers.
- Standards for agent identity. NIST’s National Cybersecurity Center of Excellence published a concept paper on software and AI agent identity and authorization on February 5, 2026, with comments due April 2, 2026. It covers identification, authorization, auditing and non-repudiation of agents. The U.S. government is treating agents as identities that need their own controls.
- Vendor consolidation. Palo Alto Networks closed its acquisition of CyberArk on February 11, 2026, per its Form 8-K, and SailPoint absorbed Entro in June 2026. NHI discovery is becoming a feature of identity platforms rather than a standalone product category.
- Agents inside every application. Gartner predicts that 40% of enterprise applications will integrate task-specific AI agents by the end of 2026, up from less than 5% in 2025, according to its August 26, 2025 press release. Each embedded agent that calls tools or data needs a credential, so the identity count follows the application count.
- The math of growth. If Palo Alto Networks’ respondents are right, machine identities grow faster than humans for another year, toward 123.7 per human. Secrets leaked in 2022 would still be 58.5% valid in January 2027. Certificate validity limits keep tightening on a published schedule, tracked on our machine identity statistics page.
Marcus Chen: For 2027 planning, watch the Discover and Retire rows of NILGI. If the next CSA survey shows unknown-agent discovery staying above 80% while agent counts grow 85%, the inventory problem is getting worse. The same is true if formal decommissioning stays near one in five. Agent counts aren’t the measure of a mature NHI program. Agents retired on purpose are.
Methodology
Collection. Axis Intelligence Research opened every source document cited here between November 6 and 7, 2026, and logged each figure with its URL and retrieval date in the accompanying CSV. Sources are the organizations that produced the measurement: survey publishers (Cloud Security Alliance, Palo Alto Networks, SANS Institute), telemetry publishers (Datadog, GitGuardian, Entro Security Labs), breach researchers (Verizon, IBM with the Ponemon Institute), incident responders (Google Threat Intelligence Group), standards bodies (OWASP, NIST), analyst forecasts (Gartner) and regulatory filings (SEC). Vendor-commissioned surveys are identified by sponsor wherever they appear.
Survey vs. telemetry. Survey figures describe the share of organizations reporting a condition. Telemetry figures describe the share of credentials or identities observed by one vendor’s platform. This page never combines the two kinds in a single calculation.
NILGI. The Non-Human Identity Lifecycle Gap Index takes five organization-level failure rates, one per lifecycle stage (discover, own, govern, respond, retire), from two Cloud Security Alliance surveys of IT and security professionals (n=383, fieldwork August–September 2025; n=418, fieldwork January 2026). Stages carry equal weight. NILGI = mean of the five rates, on a 0–100 scale where higher means a wider gap. The Retire input is 100 minus the 21% of organizations with formal decommissioning. The November 2026 figure is the baseline reading. Future readings will substitute newer CSA editions stage by stage, and any change to stages or weights will be stated on this page. NILGI measures process gaps reported by organizations. It does not count identities or weigh the business impact of a gap.
Leaked-secret half-life. Axis uses GitGuardian’s two published retests of the same 2022 cohort (70% valid in January 2025, 64% in January 2026). Annual survival = 64 ÷ 70; half-life = ln(0.5) ÷ ln(64/70) = 7.7 years. The projection assumes the 2025 revocation rate stays constant. GitGuardian describes both readings with approximate wording (“nearly 70%”, “above 64%”), so the half-life is reported to one decimal and labeled an estimate.
Visibility confidence gap. Inclusion–exclusion on two answers from the same CSA sample (68% strong visibility, 82% found unknown agents): minimum overlap = 68 + 82 − 100 = 50% of respondents, or 50 ÷ 68 = 73.5% of the confident group.
Agent containment gap. Difference between two answers in the same Palo Alto Networks survey: 91% believe they can rapidly contain a compromised AI agent; 37% apply credential revocation to autonomous agents (n=2,907). Gap = 91 − 37 = 54 points.
NHI lifecycle management among breached organizations. IBM reports 46% of breached organizations secure NHIs in AI workflows, and 55% of those use machine identity inventory and lifecycle management. Share of all breached organizations = 0.46 × 0.55 = 25.3%.
2027 ratio projection. 2026 ratios from Palo Alto Networks scaled by the same respondents’ expected 12-month growth for machine (77%), AI agent (85%) and human (56%) identities.
About This Dataset
The dataset non-human-identity-statistics.csv contains 160 rows in long format. Each row is one observation with metric, value, unit, as_of_date, geography and segment fields, plus provenance fields source_org, source_document, source_url, retrieved_date, is_primary and axis_calculated. Every Axis-derived row carries a method_note giving its formula. It covers observations from 2022 to 2026 and projections to 2028 from 13 source organizations. The file is updated when a source publishes a new edition, not on a calendar. License: CC BY 4.0. Required attribution: “Axis Intelligence Research, Non-Human Identity Statistics 2026–2027, 2026.”
Cite This Research
- APA: Axis Intelligence Research, & Chen, M. (2026, November 7). Non-human identity statistics 2026–2027: Where the NHI lifecycle breaks. Axis Intelligence. https://axis-intelligence.com/non-human-identity-statistics/
- MLA: Axis Intelligence Research, and Marcus Chen. “Non-Human Identity Statistics 2026–2027: Where the NHI Lifecycle Breaks.” Axis Intelligence, 7 Nov. 2026, axis-intelligence.com/non-human-identity-statistics/.
- Chicago: Axis Intelligence Research and Marcus Chen. “Non-Human Identity Statistics 2026–2027: Where the NHI Lifecycle Breaks.” Axis Intelligence, November 7, 2026. https://axis-intelligence.com/non-human-identity-statistics/.
What Security Teams Ask About Non-Human Identities
Is an AI agent a non-human identity?
Yes, whenever it authenticates to a system with its own credential, such as a token, key or workload identity. Palo Alto Networks counts AI agents inside its machine identity total: 79 of the 109 machine identities per human in 2026. NIST’s NCCoE concept paper of February 5, 2026 treats agent identification and authorization as an identity problem in its own right.
Should a board deck use 109:1 or 144:1 for the NHI-to-human ratio?
Use 109:1 to describe a typical enterprise, and cite it as a 2026 survey of 2,930 security leaders by Palo Alto Networks. Use 144:1 only as an example of what discovery tooling finds once deployed (Entro Security Labs, H1 2025 telemetry). The two measure different things, so they should not be averaged.
If a leaked API key is removed from the repository, is the risk gone?
No. Deleting the commit does not invalidate the credential. GitGuardian found 64% of secrets confirmed valid in 2022 still working in January 2026. Axis Intelligence Research estimates a leaked-secret half-life of about 7.7 years at the current revocation pace. Only revoking or rotating the credential removes the risk.
Who should own a service account when the person who created it leaves?
A named team, not a person, recorded at creation time. In the Cloud Security Alliance’s January 2026 survey, 51% of organizations reported no clear ownership or accountability for NHIs. Entro Security Labs found 91% of former employees’ tokens still active, so offboarding has to include the identities the employee created, not only their own login.
How fast should an exposed machine credential be revoked?
Within hours, and at the latest the same day. The Cloud Security Alliance found 24% of organizations need more than 24 hours to rotate or revoke a credential after potential exposure, and 30% take more than a day just to triage a high-severity leak. Palo Alto Networks found only 41% of organizations include revoking tokens, API keys or certificates in their incident response procedures. In the Salesloft Drift campaign, data theft ran for at least 10 days (August 8–18, 2025) before all tokens were revoked on August 20.
Why do stolen credentials rank behind exploits in the 2026 DBIR if NHIs are multiplying?
Because the DBIR codes the initial access step, where exploitation led at 31% against 13% for credential abuse. NHIs are more often abused after entry, during lateral movement and data access, and through vendors. Third-party involvement reached 48% of breaches in the same report.
Do most organizations rotate NHI credentials every 90 days?
No. The 2026 SANS State of Identity Threats & Defenses Survey found that 92% of organizations fail to rotate the majority of their NHI credentials on a 90-day cycle, and only 8% rotate more than 75% of them that often. Datadog’s telemetry agrees: 59% of AWS IAM users had an access key older than one year in 2025.
Is the “97% of NHIs are over-privileged” statistic credible?
It is a real Entro Security Labs finding (2025), but it measures any unused permission, so it describes hygiene rather than exploitable risk. A narrower telemetry measure from Datadog puts AWS EC2 workloads with risky permissions at 19.4%. For prioritization, use Entro’s finding that 1 in 20 AWS machine identities hold full admin rights.
How many organizations can retire AI agents in a controlled way?
About one in five. The Cloud Security Alliance’s April 2026 survey found 21% have formal decommissioning processes for AI agents. Retirement is the second-weakest stage in the Axis NILGI, after discovery.