Identity observability is the ability to reconstruct what an identity did, where, and on whose behalf, using events from the systems that issue and accept credentials. It answers one practical question with evidence: is this account behaving the way its owner and purpose say it should?
Most identity telemetry still describes people signing in. An analyst can usually say when an employee authenticated, from which device, and whether MFA passed. The same logs say much less about the service account that ran all night, or the AI agent a product manager connected to the CRM last week.
That gap is widening. AI agents act continuously and at machine speed, so a baseline built around working hours and daily sign-in counts has nothing to compare them against. An agent's normal is the task it was built for, and the first problem is collecting the events that describe that task at all.
What Is Identity Observability?
Identity observability means collecting identity events, from authentications to agent actions, with enough context to answer questions nobody wrote a rule for in advance. Monitoring watches for known conditions; observability lets a team ask new questions of identity activity after the fact.
Three properties separate an observable environment from a monitored one. The events cover more than sign-ins. Each event resolves to an identity with a named owner. And events from different systems share keys, such as a person record or a session ID, so they can be joined.
A quick test: mid-investigation, an analyst asks which identities used a particular OAuth grant last month and who is accountable for each. If the data answers in minutes, the environment is observable. If it takes three teams and a spreadsheet, it is monitored at best. Visibility, a related concern, covers which identities exist and what they can reach; identity observability records what they did over time.
Why Continuous Identity Monitoring Matters for Security
Authentication happens once, but access lasts. A browser session can run for a full workday, refresh tokens can outlive it, and an agent may hold a delegated grant for weeks. Most of what an identity does happens after the sign-in check that security teams log most carefully.
That is the reasoning behind continuous verification: trust granted at sign-in gets re-evaluated as conditions change. Re-evaluation depends on fresh evidence. If the only identity data a team holds is a successful authentication at 8:52 a.m., there is nothing to re-evaluate at 2 p.m.
The Australian Signals Directorate's ACSC, CISA, the FBI, the NSA, and international partners made the same point in August 2024. Their joint guide, Best Practices for Event Logging and Threat Detection, lists "an account accessing services that it does not usually access" as anomalous behavior. Both that example and "concurrent sign-ins from multiple geographic locations" describe activity after a successful login, which is where identity observability earns its keep.
Continuous does not mean storing everything forever. Retention for each log type depends on regulation, risk, and the organization's control framework. Coverage comes first: the events that describe post-authentication behavior have to exist before anyone can watch them.
Key Identity Signals to Monitor Across Your Environment
The most useful identity signals record a change to an identity, its credentials, or its reach. Sign-in counts matter less than the moment an account gains a new way in or a new permission. The table pairs each signal with where it is usually logged and what a suspicious instance looks like.
| Signal | Where it is logged | Suspicious example | ATT&CK reference |
|---|---|---|---|
| MFA challenges | IdP authentication log | A burst of push prompts to one user, then an approval | T1621 |
| Authenticator registration | IdP audit log | New MFA device added minutes after a password reset | T1098.005 |
| Credential resets | Help desk tickets and IdP audit log | Phone-requested reset for a user who is on leave | T1684.001 when obtained through impersonation |
| Session and token use | IdP, application, and proxy logs | One session cookie presented from two networks | T1550.004 |
| Role and permission grants | Cloud control plane audit log | Admin role assigned outside any change ticket | T1098.003 |
| Non-human sign-ins | Directory and IdP sign-in logs | A service account signing in interactively | None specific |
| Agent tool and API calls | Agent gateway and application API logs | An agent calling a tool it has never used | None specific |
Token signals deserve particular attention because they bypass the checks teams trust most. With token theft, an attacker who copies a valid session cookie can skip the password and the MFA prompt, because the session is already authenticated. The IdP recorded one clean sign-in, and every later request looks like the legitimate user unless the application logs the network and device each token arrives from.
Agent calls need a different kind of record. A useful entry names the agent, the person who started the task, the authority it acted under, and the resource it touched. That chain, sometimes called operational lineage for agents, is what lets an analyst tell a sanctioned task from a hijacked one.
Detecting Anomalous Identity Activity and Potential Threats
Single events rarely prove abuse. A password reset is routine, and so is a new MFA device. A reset, then a new device, then a sign-in from an unfamiliar country inside an hour is a pattern.
Push bombing that ends in an approval. The Scattered Spider advisory from CISA, the FBI, and partners (AA23-320A, revised July 29, 2025) describes actors who "sent repeated MFA notification prompts". Employees eventually pressed "Accept". The detection counts prompts per user in a short window and fires when a successful sign-in follows. These MFA fatigue attacks map to ATT&CK T1621.
A help desk reset followed by a new device. The same advisory describes actors posing as employees to get help desk staff to "reset the employee's password, and transfer the employee's MFA to a device they control". The trail runs from ticket to reset to authenticator registration (ATT&CK T1098.005) to a sign-in from a new location.
A service account that starts acting like a person. A deployment account that only ever authenticates from one build server completes an interactive browser sign-in from a laptop subnet. Nothing failed, so a failed-login rule stays silent.
One token, two networks. A session issued on the corporate network reappears minutes later from a hosting provider's address range. The logging guide's examples of anomalies include "impossible travel" and concurrent sign-ins from multiple locations, which is how this pattern usually surfaces.
An agent reaching for a new tool while its owner is offline. A reporting agent that has only read from the data warehouse calls a file-sharing API at 3 a.m. It may be a new, approved task. Without the owner and the task attached to the event, nobody can tell.
A role grant outside the change window, then data access through it. A contributor account gains an owner role on a storage account with no matching ticket (ATT&CK T1098.003). An hour later it reads data through that role, which is privilege escalation documented step by step.
These detections combine rules for known sequences with anomaly detection methods that compare an identity against its own history. The logging guide warns that anomalies "could be legitimate behaviour" and need investigation. A workable rule: alert at once when a sequence ends in a credential or privilege change that no ticket or owner explains. Route a single deviation with a plausible business reason to investigation.
Where does this end and ITDR begin? ITDR solutions apply detections and response actions to identity data, so they are one way to act on what observability collects. The two overlap heavily, and an ITDR tool is only as good as the events it receives.
Correlating Identity Signals Across Applications, Cloud Environments, and Infrastructure
Many security teams answer the observability question with "our SIEM already ingests IdP logs." Those logs show authentication. They often miss what happens after the token is issued, actions in applications outside SSO, and agent tool calls. They rarely say who owns a non-human identity.
Consider a common path. An attacker signs in through the IdP as a developer and opens the cloud console with that session. There they create an access key, which they later use from a rented server. Three systems log the activity under three identifiers: a user principal name, a role session name, and an access key ID.
Viewed separately, each event looks ordinary. Joined, they show lateral movement from a workforce identity into cloud infrastructure, plus persistence through a credential nobody issued on purpose. Identity observability depends on that join. The join needs keys: a person record that maps accounts to one human, session IDs that travel between systems, and synchronized clocks.
Source coverage comes first. The ASD and CISA guide tells cloud environments to log "all control plane operations, including API calls and end user logins". Those logs should capture "read and write activities, administrative changes, and authentication events" and flow to a central facility a SIEM can draw from.
Common Identity Observability Challenges Security Teams Face
The obstacles are rarely about detection logic. They sit in the data underneath it.
Identifiers that do not line up. One engineer can be jsmith in Active Directory, john.smith@ in the IdP, jsmith-adm for privileged work, and a personal handle in GitHub. Until those resolve to one person, a sequence spread across them reads as four unrelated users.
Non-human identities with no owner. An alert on a service account sits in a queue when nobody knows who created the account or why. The detection worked; the routing failed.
Activity outside SSO. Local accounts, long-lived API keys, and unfederated applications generate events the identity team does not see. Some produce no identity logs at all.
Agents that act as their user. When an agent works through a user's own OAuth grant, the application log typically records the human as the actor. Often only a client ID shows that software made the call. Identity observability then attributes machine-speed activity to a person who was asleep.
Volume and protection. Data-access logs are large, and retention costs force trade-offs per source. The ASD and CISA guide also warns that some actors "modify or delete local system event logs to avoid detection".
Best Practices for Implementing Continuous Identity Observability
Continuous identity observability is built source by source, and the order matters more than the tooling. These practices assume a team already runs a SIEM and an IdP.
- 1.Onboard change events first. Start with the logs that record authenticator registrations, credential resets, and privilege grants in the IdP and cloud control planes. A useful rule: if budget covers only one new source this quarter, pick the one that logs changes to identities the team cannot yet attribute.
- 2.Assign an owner before writing the detection. A detection on an identity with no owner produces an alert nobody can close. Record a named owner and a stated purpose for each service account and agent.
- 3.Give agents their own identities. An agent acting under its own identity, with delegated authority recorded per call, produces logs an analyst can attribute. One acting through its user's grant does not.
- 4.Centralize and protect the record. Follow established audit logging practices: central collection, restricted delete rights, and storage segmented from the systems being logged.
- 5.Feed findings back into access decisions. A signal is more valuable when it changes the next authorization, such as stepping up authentication after a new device registration. An adaptive access policy is where that feedback lands.
- 6.Rehearse against real sequences. Walk the help desk scenario from AA23-320A through the team's own logs. Each step that leaves no event is a gap to close before an attacker finds it.
Each practice leads back to the same problem. Most identity signals reach the SOC with no owner or purpose attached. A detection that cannot say whose identity acted, and on whose authority, waits in a queue. NewCore's view is that identity is the attack surface: the actors in AA23-320A bought or reset real credentials and then registered their own MFA tokens. The useful question is whether an identity is behaving as its owner would.
Answering it takes identity context the logs do not carry on their own. NewCore's Discover capability resolves a person's accounts across identity providers, directories, and HR systems into one identity record. It links each AI agent to its owner and traces access paths, including outside SSO. That mapping is what an analyst needs to read a sequence spread across four usernames as one person's activity.
For agents, NewCore Govern captures tool calls, sessions, and approvals in one audit trail and traces each automated call to the person accountable for it. It treats denied access attempts as a primary threat signal, not background noise. Identity Foundation keeps authenticators in one inventory with strength scores, owners, and last use, and can suspend a risky one with a logged reason. None of this replaces the SIEM or the ITDR tooling, and a team can start alongside its current identity platform. NewCore feeds that stack the context that makes a signal attributable. A detection without an owner is only a question.
Your identities are already acting. Make every action answer to someone. Request a Demo →
Frequently Asked Questions
What is the difference between identity observability and ITDR?
They overlap. Identity observability is the collection and correlation of identity events with owner and purpose attached. ITDR is a category of tools that detects and responds to identity abuse using that kind of data. A team can run ITDR with thin telemetry, but its detections will only be as complete as the events it receives.
Is identity observability the same as identity monitoring?
Not quite. Monitoring checks identity activity against conditions defined in advance, such as failed-login thresholds. Observability keeps enough context to answer questions nobody planned for, such as which identities used a specific grant last month and who owns them. Most mature programs use both.
Which logs are needed for identity observability?
It depends on the environment, but the usual starting set is IdP authentication and audit logs, help desk reset records, cloud control plane logs, and application logs for systems outside SSO. Agent gateway or API logs matter once AI agents act on company systems. Change events tend to deliver value before sign-in volume does.
How do you observe AI agent activity?
Give each agent its own identity and log every tool or API call with the agent, the person who started the task, the authority used, and the resource touched. When agents act through a user's own credentials, most logs attribute their actions to that user, which makes misuse hard to separate from normal work.
Can a SIEM provide identity observability on its own?
Sometimes, if the right sources feed it and identities resolve to owners. A SIEM can store and correlate identity events well. It cannot create events that were never logged, and it cannot tell who owns a service account unless another system supplies that context.
Discover every agent acting in your business.
Book a DemoIdentity Security, NewCore
The team behind NewCore's Identity 101 guides, covering agent security, identity security, workforce security, identity governance, and identity discovery. These guides draw on NewCore's own product and security work, translated into practical, vendor-neutral guidance for security and IT teams managing identity across people, software, and AI agents.
View all articlesCite this guide
The NewCore Team. (2026, November 18). Identity observability: tracking identity activity and detecting emerging threats. NewCore. https://newcore.com/identity-101/identity-observability


