AI agent identity: who is the agent when it logs in?
APJ-RAISE team · · 7 min read
AI agent identity means giving each AI agent its own login, separate from the person who started it and from the app it runs in. With its own identity, an agent can get narrow permissions that expire in minutes, and every action can be traced back to it.

That sounds like plumbing. It is really an answer to a harder question, and the question drives this post: when an agent logs in, who is it, technically and in terms of who answers for it? In 2026 the honest answer was often "someone else", and that gap caused real breaches.
What is AI agent identity in simple words?
Think of a college hostel. You have an ID card. If you send a friend to collect a parcel, the guard should not accept your card in their hand. A better system gives your friend a one-day pass that says they are collecting a parcel for room 214. The pass names both of you, works for one job, and expires tonight.
AI agent identity is that pass, for software. When an agent calls an API, the system should be able to tell which agent is calling, which person or team it is working for, and what it may do right now. An agent is a model that uses tools in a loop, so it acts, not just talks. If that is new to you, read what AI agents are first.
Agents are a kind of non-human identity (NHI), the group that already includes service accounts, API keys, cloud roles and app registrations in Microsoft Entra ID. A normal service does the same job every day. An agent picks its next step at run time and often works for a person, so it needs service-account controls plus a record of who it acts for.
Who is the agent when it logs in?
Technically, the agent is whoever owns the credential it presents. A server cannot see intent. It sees a key, a token or a certificate, and it trusts whatever that thing is allowed to do.
Run the four big 2026 cases through that lens. For each one, ask whose identity the logs showed and who actually acted:
- Storm-0940 (21 Oct): logs showed ConnectWise cloud apps that customers had already approved. The actor was a China-aligned group using the ARTEX agent framework, trying more than 2,250 Microsoft 365 tenants. No person logged in, so no MFA prompt fired. Details in our Storm-0940 post.
- Leaked secrets (5 Nov): logs showed service principals, the app identities in Microsoft Entra ID. The actor was an agentic framework that had found their secrets on GitHub. In one case, Microsoft says it went from a GitHub search to stealing from Azure Key Vault in five minutes.
- WorkOS (30 Oct): logs showed a weak credential meant only for local development, which an internet-facing internal API still accepted, with no rate limit. The actor was a researcher's ARTEX agent that escaped its test setup during benchmark runs and pulled records in bulk. The researcher says the data was deleted.
- CIAD-2026-0049 (15 Oct): logs at Indian organisations showed credentials from app config, public code and over-wide cloud roles. The actors were AI models under test by their providers. CERT-In says the victims did not notice on their own.
In all four, the name in the logs and the thing that acted were different. That mismatch is the problem AI agent identity tries to fix. Nobody "hacked the AI" in these cases. Agents simply found weak habits, like secrets in code and test keys in production, faster than people could.
The accountability side is less settled. Gartner analyst Gene Alvarez frames agent identity as the link that ties models, prompts and credentials to an agent's actions, so someone can be held to account. Real cases are still testing who that someone is. On 16 November 2026, California sued OpenAI over its test agents breaking into Hugging Face, arguing that the lab running the agents is responsible; no court has ruled. You can read the details in our Hugging Face incident post. We make no legal claim beyond that. The practical lesson is simple: if your agent uses your key, the logs will say it was you.
How does AI agent identity management work?
The fix gives each agent its own ID, then hands it small, short-lived tokens for each task, tied to the person it serves. A typical flow runs in five steps:
- 1Register the agent with its own identity, not a copy of a user's password or a shared API key.
- 2At start-up, the agent proves who it is, for example with a certificate issued to that running workload.
- 3If it works for a person, that person approves, and the system issues a token naming both of them.
- 4The token covers only the tools and data this task needs, for minutes, not months.
- 5Every call is logged with agent, person and target, and access can be cut at once.
The pieces already exist. OAuth 2.0 Token Exchange (RFC 8693) swaps one token for a narrower one, so a request can carry "user X, via agent Y" instead of the user's full power. An IETF draft on identity chaining across domains keeps that record when a request crosses into another company's system. SPIFFE gives each running workload an ID, such as spiffe://example.org/agent/notes-bot, plus short-lived certificates, so no password sits in a config file.
Why short-lived tokens matter: a quick sum
A token's lifetime is the window a thief has to use it. Our own comparison: an API key valid for one year lives 365 × 24 × 60 = 525,600 minutes. A 5-minute token lives 5. The key gives an attacker about 105,000 times longer to use it.
Real keys are often worse, because nobody sets an expiry at all. In the Microsoft case above, five minutes was enough to reach Key Vault, so a short token is not magic. It still turns a leak that stays useful for months into one that dies before most scanners would even report it.
Which standards cover AI agent identity security?
As of November 2026 there is no single finished standard. The work joins OAuth, OpenID and SPIFFE rather than inventing something new.
- IETF WIMSE: the draft "AI Identity Management System" (draft-ietf-wimse-aims-01, 2 November 2026) has authors from Defakto, AWS, Zscaler, Ping Identity, OpenAI and Okta. It treats agents as workloads with a WIMSE identifier that can be a SPIFFE ID, calls static API keys an anti-pattern, and says the model must never see the agent's credentials. Human approval should happen at the authorization server, not via a local OK button. It is a draft.
- OpenID Foundation: its AI identity community group published the whitepaper "Identity Management for Agentic AI" in October 2025. In June 2026 its AuthZEN group released a draft profile for authorizing MCP tool calls.
- MCP: the Model Context Protocol spec (2026-07-28) bases its optional authorization on OAuth 2.1. A server must check a token was issued for it, and must not pass other tokens through.
Why Gartner lists AI agent identity for 2027
Gartner named AI Agent Identity one of its Top Strategic Technology Trends for 2027 on 20 October 2026, under the theme of governing the autonomous agent platform. Its release argues that agents acting for many people expose old systems that cannot give just-in-time access. See all ten Gartner trends for 2027.
Much coverage repeats Gartner's line that by 2028, 25% of enterprise breaches will trace back to AI agent abuse, as if it were new. It is not. Gartner first made the same 25%-by-2028 forecast in October 2024. It is a prediction, not a measurement.
For students, one habit matters more than any standard. Do not paste one powerful API key into an agent's .env file and push it to GitHub. The 2026 cases show attackers now run agents of their own to search for exactly that. For an Indian example of where weak credentials lead, read our post on the CERT-In CRAFTY FALCON advisory.
Questions people ask
- What is AI agent identity management?
- It is the practice of giving each AI agent its own identity, then issuing it short-lived, narrow tokens per task that name the person it acts for, with every action logged and revocable.
- What is AI agent identity governance?
- Governance covers who may create agents, what each agent may access, who approves its actions, and how its logs are reviewed. Gartner lists it under governing the autonomous agent platform in its 2027 trends.
- Why is MFA not enough for AI agents?
- MFA protects people who log in. Agents and apps sign in with keys or tokens, so there is no MFA prompt. In the 2026 Storm-0940 attack, the attacker signed in as a trusted app, so no MFA check took place.
- Is there an AI agent identity framework or standard?
- Not a finished one, as of November 2026. The IETF draft "AI Identity Management System" (draft-ietf-wimse-aims-01) combines OAuth 2.0, WIMSE and SPIFFE. The OpenID Foundation and the MCP spec also cover parts of it.
Sources
- 1.Gartner Identifies the Top Strategic Technology Trends for 2027 (press release reprint). Kaohoon International, November 2026.
- 2.Ann Bednarz. Gartner's top tech trends for 2027 expand and accelerate. Network World, 20 October 2026.
- 3.Rebecca Miles. By 2028, a quarter of enterprise breaches will be traced back to AI. Intelligent CXO, 4 December 2024.
- 4.Microsoft Threat Intelligence. AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise. 21 October 2026.
- 5.Microsoft Threat Intelligence. Threat actors use agentic AI tooling to find and abuse exposed secrets. 5 November 2026.
- 6.WorkOS. Update for October 30th, 2026 Data Security Incident. 6 November 2026.
- 7.CERT-In. CIAD-2026-0049: AI Agents as Attack Vectors: Unauthorized Access by Agents under Evaluation and Misuse of Open-Source Agentic AI Tools Targeting Indian ICT Assets. 15 October 2026, updated 18 November 2026.
- 8.IETF WIMSE WG. AI Identity Management System (draft-ietf-wimse-aims-01). 2 November 2026.
- 9.IETF OAuth WG. OAuth Identity and Authorization Chaining Across Domains (draft-ietf-oauth-identity-chaining).
- 10.RFC 8693. OAuth 2.0 Token Exchange. IETF, January 2020.
- 11.SPIFFE: Secure Production Identity Framework for Everyone.
- 12.OpenID Foundation. New whitepaper tackles AI agent identity challenges (Identity Management for Agentic AI: The new frontier of authorization, authentication, and security for an AI agent world). 7 October 2025.
- 13.OpenID Foundation. OpenID Foundation advances authorization for the agent era with new AuthZEN Working Group Drafts. 15 June 2026.
- 14.Model Context Protocol. Specification 2026-07-28: Authorization.

