Discover what post-authorization security means for the AI agent era, the three questions it answers, and practical steps for auditing agent data access.

Look at the SaaS breaches that made headlines over the past year and you'll notice something odd. Almost none of them broke authentication.
In the Salesloft Drift incident, attackers stole OAuth refresh tokens that customers had issued to a chatbot integration. For about ten days in August 2025 they used those tokens to query Salesforce and other systems at more than 700 organizations. The queries came from a trusted, pre-approved app, so they looked like normal chatbot traffic. Nobody's login controls were bypassed. The tokens were real.
The Vercel breach in April 2026 followed the same shape. An employee at Context.ai, an AI tool vendor, picked up infostealer malware. The attackers used that foothold to get at OAuth tokens, including one a Vercel employee had issued to Context.ai with broad "Allow All" access to their Google Workspace account. From there they walked into Vercel's internal systems. Again, every credential involved was valid.
I keep coming back to these two cases because they show where security effort actually goes. We spend enormous energy on the moment of access: MFA, SSO, conditional access, consent screens. Then a token gets issued and we mostly stop watching. That gap has a name, or at least it should. We call it post-authorization security.
Post-authorization security refers to the ongoing security controls and monitoring applied after a user has been granted access to a system. It ensures continuous compliance, detects post-login anomalies, and prevents unauthorized lateral movement by enforcing zero-trust policies throughout the session.
Most access controls answer a yes or no question before anything happens. Can this identity reach this resource? Post-authorization security asks a different question after the fact, and keeps asking it: given that this identity could reach the data, should it have, in this way, at this volume, for this purpose?
A simple way to hold the difference between the Authorization and Post-authorization security layers:
• Authorization decides what an identity is allowed to do.
• Post-authorization checks what an identity actually did with that permission, and whether that matches who it is and why it was given access, then closes the gaps before they can be used.
You might ask whether existing tools already cover this. User behavior analytics and insider risk programs have watched user activity for years, but they aren't enough. Anomaly detection looks for behavior that is unusual, and illegitimate access usually looks normal. The Drift tokens produced traffic that looked like the chatbot doing its job. An agent running on a departed employee's grant keeps the same schedule it always had. A sales rep reading payroll files at normal volume and normal hours raises no statistical flag at all. None of these is an anomaly.
Each is wrong because of who is acting, what the data is, and why the access was granted, and only that context can tell you so. That gets harder as most of the access to enterprise data stops coming from humans at a keyboard.
Every access moves through three distinct moments: Pre-Authorization, Authorization, and Post-Authorization.
1. Pre-Authorization (Guarded)
Asks: Who should be allowed?
Access is decided in advance through entitlements, roles, data classification, and security posture. This stage is owned and covered by IAM, IGA, PAM, and DSPM.
2. Authorization (Guarded)
Asks: Is this request permitted now?
This is the runtime grant-or-deny evaluation at the gate based on policy rules. This stage is owned by policy engines and DLP rules.
3. Post-Authorization (Blind Spot)
Asks: Should this access have happened?
Evaluates whether the permitted access was legitimate by verifying if it was the right identity, intent, and data. Authorization cannot stop at the grant—the moment the access actually happens is the one nothing re-checks. This crucial layer is currently owned by no one yet, and is the exact gap Dawn owns.

An AI agent is software that acts with a person's delegated intent but has no person at the keyboard. That combination breaks a lot of assumptions our tools quietly rely on.
When an agent reads a file, the log often records the wrong actor. Sometimes it's the human who connected the agent weeks ago. Sometimes it's a shared service account that a dozen workflows use. Sometimes it's an account that was created for the agent and looks exactly like an employee. And sometimes nothing useful gets logged at all.
The pattern we see again and again is that agent access is allowed but unattributable. The permission check passed, so nothing looks wrong. But if you ask who actually touched that board deck, on whose behalf, and why, there is no single record that answers it.
That matters for two reasons. The first is audit: you can't govern what you can't attribute. The second is attack. An agent that has been manipulated through a poisoned document or a malicious tool doesn't throw an error. It keeps working, using access it legitimately holds, and it looks like a system doing its job. If you only watch the front door, a compromised agent and a healthy one are indistinguishable.
The major identity vendors see this too. Okta and Microsoft have both shipped dedicated identity for AI agents in 2026, which is real progress on the issuing side. But even the people building those standards point out that recording the full chain of who asked which agent to do what is still left to the company deploying the agents. Issuing a good credential is not the same as knowing how it was used.
If you strip it down, a security team needs three answers about the agents in its environment:
1. Which AI agents are touching our data?
2. What sensitive data are they touching?
3. Was each access legitimate, given who the agent acts for and what it was meant to do?
Answering the third one well means rebuilding a chain, not reading a single log line: the identity, its intent, the agent and any sub-agents, the MCP server or skill it used, the tool it called, and the data it reached, with that data's classification attached. Permission covers one link of that chain. Legitimacy depends on all of it.
You don't need a new platform to take the first steps. A few things any team can do this quarter:
• Inventory your OAuth grants. List every third-party and AI app with access to Google Workspace, Microsoft 365 and your CRM, with its scopes. Both breaches above started with a grant someone approved and forgot.
• Find the agents that look like people. Ask teams whether they've created user accounts for bots or agents. Those accounts are invisible in most logs.
• Tie access to data classification. An agent reading a lunch menu and an agent reading payroll are not the same event. Your monitoring should know the difference.
• Put agents in access reviews. If your quarterly reviews only cover humans, they certify a shrinking share of what can reach your data.
Dawn is built for this layer. We connect to the places where your data lives, like Google Workspace and Microsoft 365, and to the platforms your agents run on. Then we stitch those scattered records into one story per access.
For every agent touching your data, Dawn shows who it acts for, what it reached, how sensitive that data is, and whether the access fits. We do this with dynamic access and not through policies post authorization. Agents are discovered automatically so nothing reaching your data stays invisible. When something doesn't fit, such as an agent still running on a departed employee's grant or an unowned agent reading regulated data, Dawn raises a finding with the evidence and the recommended fix, and routes it to the team that owns it.
If you want to see what your agents are actually doing with your data, get in touch. To learn more, explore our product page.