Blog · · 3 min read
A password reset is not a session revocation
Resetting a compromised password feels like containment. If existing sessions and tokens survive the reset, the attacker keeps their access while the incident is recorded as closed.
Red Cell · identity · SSO · incident response
When an account is suspected compromised, the first move in most runbooks is to reset the password. It is fast, it is visible, and it produces a clear line in the incident record. The problem is what it leaves untouched. An attacker who already holds a valid session does not need the password any more, and in many single sign-on setups a password change does nothing to the sessions and tokens that were issued before it.
Authentication happens once; sessions last
A password is checked at sign-in. After that, access rides on things issued as a result: a session cookie with the identity provider, session cookies with each application the user opened, and often access and refresh tokens used by APIs and mobile clients. Each of those has its own lifetime and its own rules about when it stops working.
Changing the password changes what will be accepted at the next sign-in. Whether it also ends what is already issued depends on configuration at several layers — the identity provider, each connected application, and any token-based integrations. It is common for one of those layers to keep honouring an old session until it expires on its own schedule.
Why this turns containment into a gap
Consider the sequence from the defender’s side. A suspicious sign-in is noticed, the password is reset, the user signs in again, and the ticket is closed. From the attacker’s side, a stolen session cookie or refresh token continues to work, because nothing revoked it. The response is recorded as successful while the access it was meant to end is still live.
This is not hypothetical for us. In a recent engagement, tokens issued by the client’s SSO provider stayed valid through a password reset. A stolen session therefore survived the exact incident-response step intended to end it.
What to check in your own environment
You can test this without an attacker. Use a test account and two browsers or devices:
- Sign in on both, and open a few connected applications on each.
- Reset the password from the first.
- On the second, keep using the applications that were already open, and note which ones keep working and for how long.
- Repeat for any mobile apps, API clients or integrations that use refresh tokens, since those often follow different rules from the browser.
The answer you want is that every existing session ends promptly. The answer you often get is a mix: the identity provider ends its own session, some applications follow, and others carry on until their own session times out.
Make revocation an explicit step
The fix is partly configuration and partly process. On the configuration side, most identity providers can revoke a user’s sessions and refresh tokens directly, and many applications can be set to re-check the user’s status rather than trusting a long-lived local session. Where an application cannot do either, shorter session lifetimes reduce how long a stolen session stays useful.
On the process side, the incident runbook should say what it means:
- Revoke all sessions and refresh tokens for the user at the identity provider.
- Revoke or rotate any application-specific tokens, keys or app passwords.
- Review and remove any MFA methods, recovery options or mail rules added during the suspect window.
- Then reset the password — and confirm, don’t assume, that old sessions are gone.
Test the response, not just the control
This class of problem is easy to miss because every individual control is working as configured. The password policy is fine, the MFA is fine, the sign-in logs are fine. The weakness only appears when you walk through what happens after a compromise, step by step, and check whether each step actually removes access.
That is why we treat incident-response steps as part of the attack surface during an engagement. A test that ends at “we obtained a session” tells you about the way in. A test that goes on to ask “would your response have ended it?” tells you whether the way out works. The services page describes how that is scoped.