Fourteen months of acquisitions rewrote the non-human identity map: CyberArk into Palo Alto Networks ($25B, closed February 11, 2026), Astrix into Cisco, Entro into SailPoint, Oasis into Cyera. The independent pure-plays left standing: Token Security, Aembit, and Clutch Security.
If you shortlisted NHI vendors in early 2025 and are only now getting budget, your shortlist is obsolete. This post is the September 2026 state of the map: who bought whom and for how much (with reported figures labeled as such), who is still independent, what the identity platforms now ship for AI agents, and the standards layer that should drive your scoring more than any logo. Last verified: September 5, 2026.
Why did AI agents break non-human identity?
The numbers were already bad before agents. CyberArk's 2025 Identity Security Landscape — a study of 2,600 security decision-makers across 20 countries — put machine identities at more than 80 to 1 against humans, with nearly half of those machine identities holding sensitive or privileged access. The Cloud Security Alliance's May 2026 whitepaper on the NHI governance vacuum reports an average of 45 to 1, reaching 144 to 1 in cloud-native environments.
Agents made the governance gap acute. In the CSA and Oasis Security "State of Non-Human Identity and AI Security" survey published January 2026, 92% of respondents said they are not confident legacy IAM can manage AI and NHI risks, and 78% have no documented policies for creating or removing AI agent identities. Meanwhile SailPoint's own survey found 82% of organizations already using AI agents. That combination — near-universal adoption, near-zero lifecycle policy — is what every deal below was priced against. We argued last year that the service-account framing was already dead; see agentic identity: non-human identity is dead.
What did the M&A wave actually change?
Between July 2025 and July 2026, most of the category's best-funded names were absorbed into platforms. The verified record:
| Acquirer | Target | Value | Status |
|---|---|---|---|
| Palo Alto Networks | CyberArk | ~$25B ($45 cash + 2.2005 PANW shares per share) | Announced July 30, 2025; closed February 11, 2026. Platform being integrated into Cortex and Strata. |
| Cisco | Astrix Security | ~$400M — reported by Israeli press; Cisco did not disclose | Intent announced by Cisco; completion reported May 4, 2026. Astrix feeds Cisco Identity Intelligence, Duo, and Secure Access. |
| Cyera | Oasis Security | ~$1B (~$700M cash, remainder in Cyera shares) | Announced July 28, 2026; completion since reported. Cyera's third acquisition of 2026. |
| SailPoint | Entro Security | ~$200M — reported; undisclosed by both companies | Announced June 15, 2026; closed June 29, 2026. Powers SailPoint's Agentic Fabric. |
| Delinea | StrongDM | Undisclosed | Announced January 15, 2026; closed March 5, 2026. Continuous just-in-time authorization for human and non-human identities. |
The pattern is stark: of the six best-known dedicated NHI startups — Astrix, Oasis, Entro, Token, Aembit, Clutch — three were acquired between May and July 2026. The Cyera–Oasis deal was called the second-largest cybersecurity deal of 2026 to date. We covered the strategic logic of the wave in NHI consolidation: what the identity acquisitions mean for AI agent security; this post is about what the board looks like after it.
The buyer's takeaway is uncomfortable but simple: for an acquired vendor, you are no longer evaluating the startup's demo — you are evaluating the acquirer's roadmap, integration timeline, and appetite to keep the product standalone. Post-close reality moves fast: Palo Alto is already cutting more than 500 CyberArk jobs, and CyberArk's press URLs now redirect to paloaltonetworks.com.
Who is still independent?
- Token Security — "machine-first" NHI security: inventory, ownership mapping, lifecycle, and exposure remediation for machine identities and AI agents. Founded 2023; $20M Series A led by Notable Capital (January 2025), $28M total. Reported triple-digit growth in 2025 on agentic-AI demand; customers include HPE, Udemy, and Elastic; RSAC 2026 Innovation Sandbox finalist. Remains independent as of publication.
- Aembit — Workload IAM: policy-based, secretless access between workloads, AI agents, and target services. Enforcement-first positioning rather than discovery-first. $25M Series A led by Acrew Capital, with Okta Ventures and CrowdStrike's Falcon Fund on the cap table. Publishes serious engineering work on MCP and OAuth 2.1 for AI authorization. No Series B announced as of publication.
- Clutch Security — "Zero Trust for NHIs": replace rotation policies with ephemeral, just-in-time credentials. $20M Series A led by SignalFire (January 2025), $28.5M total raised. Remains independent as of publication.
- GitGuardian (adjacent) — coming at NHI from secrets detection, now shipping NHI governance across cloud IAM, secrets managers, and SaaS. Its $50M Series C (February 11, 2026, led by Insight Partners) was explicitly positioned around the NHI crisis and the AI agent security gap.
The honest framing for all four: independence is a snapshot, not a guarantee. Three of their closest peers were acquired within a single quarter.
What do the platform players now offer?
- Okta has moved fastest in public. Cross App Access (XAA), introduced June 2025, extends OAuth so the enterprise identity layer — not each individual app — governs which agents may reach which applications, with early ecosystem support from Box, Boomi, Glean, and Automation Anywhere. Agent SSO went GA on August 24, 2026: XAA-capable agents become first-class identities in Okta Universal Directory with the same governance as human employees, included in core Okta SSO plans at no extra cost.
- Microsoft Entra Agent ID, announced at Build in May 2025 and now listed as generally available in Microsoft's docs, extends Entra to authenticate, authorize, and govern AI agents, with provisioning partnerships with ServiceNow and Workday and declared support for A2A and MCP identity work. August 2026 brought Entra Agent ID for Dataverse.
- SailPoint ships Harbor Pilot agents inside Identity Security Cloud and, post-Entro, an Agent Identity Security capability with connectors that discover and govern agents from Copilot, Bedrock, Vertex AI, Agentforce, ServiceNow, and Snowflake Cortex.
- CyberArk (Palo Alto Networks) made its Secure AI Agents solution generally available November 4, 2025 — each agent treated as a privileged autonomous identity, with shadow-agent discovery and privilege controls — now being folded into Palo Alto's Cortex and Strata platforms.
- Cisco (via Astrix) and Cyera (via Oasis) are integrating discovery, lifecycle, and threat detection for NHIs into zero-trust and data-security platforms respectively.
How is agent identity different from a service account?
Four ways, and they map directly to evaluation criteria.
- Credential lifecycle: ephemeral versus static. Classic service accounts run on long-lived static credentials — OWASP's NHI Top 10 lists Long-Lived Secrets (NHI7) and Improper Offboarding (NHI1) among the top risks. Agents spawn dynamically and should die with the task. Score vendors on just-in-time, short-lived credential issuance, not rotation cadence.
- Delegation chains. An agent acts on behalf of a user, and often on behalf of another agent. The base primitive is RFC 8693, OAuth 2.0 Token Exchange (IETF, January 2020), with its act and may_act claims — but OAuth was designed for single-hop delegation, and end-to-end tracking of user → agent → sub-agent → API chains is still Internet-Draft territory. Ask vendors precisely how they record who asked for what, through whom.
- Per-task authorization. The MCP authorization spec makes every MCP server an OAuth 2.1 resource server: token validation is mandatory, PKCE is mandatory, and resource indicators restrict token audience. Okta's XAA applies the same move at the enterprise layer. A vendor that cannot speak fluent MCP OAuth 2.1 is selling you last decade's service-account tooling with a new label.
- Attribution. OWASP's NHI10 — humans using non-human identities — is the classic audit failure. Agents invert it: non-human identities acting as humans. Agent-to-agent traffic adds another layer, which is what Google's A2A protocol (open-sourced April 2025, donated to the Linux Foundation that June with 100+ supporting companies) standardizes with Agent Cards and task-scoped auth. The fix in both directions is the same: agents as first-class directory identities with their own audit trail, never shared API keys.
Identity answers who the agent is and what it may touch. It does not watch what the agent actually does inside an approved session — that is runtime's job, and the two layers fail differently; we drew that boundary in AI agent runtime security. It is also where AccuroAI's agent security sits alongside your identity stack: runtime enforcement at under 38ms p99, screening 14M+ prompts daily across our customer base, with 60+ secret types detected in agent traffic — the layer that catches what a perfectly-issued token still allows.
How should you choose an NHI vendor in 2026?
- If the vendor was acquired: evaluate the acquirer's integration roadmap and get contractual protection on standalone availability and pricing through your term.
- If the vendor is independent: assume it may not stay that way; weight open-standards support (RFC 8693 token exchange, MCP OAuth 2.1, A2A, XAA) over proprietary features, because standards survive acquisitions and dashboards do not.
- If you are platform-first: Okta's Agent SSO at no extra cost and Entra Agent ID's GA mean your existing IdP may already cover baseline agent identity — test that before buying a dedicated tool, then buy dedicated tooling for what the platform demonstrably misses.
- Either way: score on delegation-chain audit, ephemeral credential issuance, and agent offboarding — the 78% of organizations with no agent lifecycle policy are the breach reports of 2027.
Frequently asked questions
Which NHI vendors are still independent in 2026?
As of September 5, 2026: Token Security, Aembit, and Clutch Security among dedicated NHI pure-plays, plus GitGuardian approaching from secrets security. Astrix (Cisco), Oasis (Cyera), and Entro (SailPoint) were all acquired between May and July 2026.
How much did the NHI acquisitions cost?
Palo Alto Networks paid ~$25B for CyberArk (verified, closed February 11, 2026). Cyera agreed to pay ~$1B for Oasis, about $700M of it cash. The Cisco–Astrix (~$400M) and SailPoint–Entro (~$200M) figures are reported by press only; neither acquirer disclosed a price.
Is an AI agent identity just a service account?
No. Service accounts are static, long-lived, and single-purpose. Agents are autonomous, ephemeral, spawn dynamically, and act on behalf of users through multi-hop delegation chains — which is why OAuth token exchange (RFC 8693), MCP's OAuth 2.1 authorization model, and A2A exist, and why legacy IGA tooling struggles with them.
What standards should I require from an agent identity vendor?
RFC 8693 token exchange for delegation, the MCP authorization spec (OAuth 2.1 with mandatory PKCE and resource indicators) for tool access, A2A for agent-to-agent communication, and XAA support if you run Okta. Purpose-built agent extensions to OAuth exist only as IETF Internet-Drafts — treat any vendor claiming to "implement the agent OAuth standard" with care.
Does identity governance replace runtime security for agents?
No — identity controls who an agent is and what it can reach; runtime security controls what it actually does with that access, in the moment. You need both layers, from the identity platform and from tooling built for agent runtime behavior respectively.