This website uses cookies

Read our Privacy policy and Terms of use for more information.

Quick one before the Wire. I made the quarterfinals of iHeartRadio's Top Host contest, and only one host in my group moves on. The vote is free, you can cast one every day, and it closes Thanksgiving night. Ten seconds, right here. More on it at the bottom.

In partnership with

Arnold Schwarzenegger. Jay Shetty. And You.

Codie Sanchez, Colin & Samir, Joanna Stern, and Oliver Darcy are here too. They all built newsletters (and podcasts and communities and websites) on beehiiv.

Some came with massive followings. Others built theirs along the way. That’s the thing about beehiiv: with so many ways to earn and grow, the platform is ready for whatever you’re building next.

-THE WIRE  THIS WEEK IN IDENTITY

N°01 · BREACH

Agents found app secrets on GitHub and signed in as the apps

Microsoft Threat Intelligence reported on November 5 that threat actors are using agentic AI tooling to search public GitHub repositories for exposed secrets, including application credentials and AI service API keys, and then act on what they find. In follow-on activity Microsoft traces back to October 28, the attacker used those credentials for service principal sign-ins on dozens of applications across multiple tenants, then pulled data from Key Vaults and storage accounts and abused Azure AI services. In one case Microsoft watched an agent go from searching GitHub to exfiltrating Key Vault data in five minutes. Todyl, which has been tracking the same infrastructure since November 1, counted 27 organizations reached by its November 3 update, many of them downstream clients of managed service providers.

This is exactly the alert this issue is about. Because these were app sign-ins, there was no MFA prompt and user-scoped Conditional Access never applied. Todyl's hunting guidance is all about workload identity behavior: one address signing in as service principals across unrelated tenants, a Resource Manager token followed within seconds by a Key Vault token, and then apps that had never been used before suddenly authenticating. Containment has a catch too. Microsoft notes that after you remove an app's secrets, tokens already issued stay valid until they expire. Your analyst can only run that hunt if they know which apps are supposed to be signing in and who owns each one.

N°02 · DISCLOSURE

An AI workflow platform lost customers' root API keys

Trigger.dev, a platform developers use to run background jobs and AI agent workflows, disclosed that on November 1 infrastructure alerts flagged anomalous activity and an attacker obtained customer root API keys for some environments. A root key gives full access to that environment's settings, including its environment variables. The company says 0.66% of customers were affected, revoked the exposed keys, and told those customers to treat the credentials in those environments as compromised and rotate database passwords, third-party API keys and access tokens. It has not disclosed the attack vector. It has since released IP allowlisting for scoped API keys and OIDC support.

An agent workflow's environment variables usually hold the credentials for every service its tasks call, so one platform key opened all of them. When a vendor sends you that email, the response depends on two things you should already have: a list of which credentials lived in that environment and an owner for each. If you only start building that list after the notice arrives, the rotation is going to be slow.

Hey {{first_name|Jedi}},

Last edition ended with a promise to take agent identity into the SOC. Here is the most useful place to start. Pick one production agent and ask your SOC to walk through a hypothetical alert involving it. Start with the event the analyst would actually receive. Can they identify the agent, find the person responsible for it, see what it was allowed to do, and locate the control that would stop its access?

That is a useful acceptance test for an agent identity program. It brings the inventory, the ownership record and the response procedure into the same conversation, and it gives you something concrete to fix before a real incident forces the handoff.

The foundations are already written down. OWASP's AI Agent Security guidance recommends recording agent decisions, tool calls and outcomes, along with security-relevant alerts. Its logging guidance also calls for event context that identifies the actor, action, timing and result. Those are sensible foundations for the handoff. The work is making them show up where the analyst looks.

Put the identity where the analyst works

Start with a small integration you can inspect. Choose an agent identifier in your inventory and document how it maps to the principals and sessions appearing in your security logs. Put the resolved identity, the accountable owner, the approved resources and the relevant execution context into the analyst's normal investigation view.

Test that mapping with an actual event from the selected agent. A field named agent_id in a schema proves very little about whether your join works. Treat an unresolved identity as an explicit gap for investigation, and preserve the source identifiers so the analyst can check the mapping instead of trusting a friendly label.

OWASP warns that event information arriving from another trust zone can be missing or forged. Keep that limitation in view when assessing an agent's own account of its actions. Use a correlation reference rather than copying access tokens or passwords into the alert.

The implementation choice here is yours. The acceptance test is straightforward: another analyst should be able to follow the identity from the alert back to the inventory without asking the agent developer to translate it.

❝

A field named agent_id in a schema proves very little about whether your join works.

Build a baseline while keeping controls active

For the selected workflow, write down the intended tools, permitted resources, expected operating conditions and the actions that require approval. Use observed activity to refine that picture, and keep the observation period and any changes to the workflow visible during triage.

Leave room for an agent's work to vary. A new destination deserves investigation in context, and a familiar destination still needs an authorization check. Keep access restrictions and security detections active while you learn what normal looks like. OWASP recommends minimum tool permissions, per-tool scoping and explicit authorization for sensitive operations, alongside monitoring for unusual behavior. Use those controls together when you design the workflow.

Prove what containment actually stops

Give the analyst a named escalation path and a response procedure for the selected agent. Document who may pause execution, who may change access and how evidence will be preserved. Include what happens when the primary owner is unavailable.

Be precise about the credential mechanism. OAuth's token revocation specification distinguishes token designs and acknowledges propagation delays. Short-lived access tokens bound an exposure window, but your system design still determines how revocation reaches the resource accepting them. Microsoft's note in this week's Wire, that issued tokens outlive a removed secret, is the same lesson from the field.

For your exercise, identify the control you would use and the protected resource where you would check its effect. Any live containment test needs an approved scope and a safe test resource. Record the observed result, including any remaining access or delay. A success message from the control plane tells you the request was accepted. Check the resource to learn whether access actually stopped.

The Last Word

Make the handoff observable. Select one agent, trace one event into the analyst's working view, and document one authorized containment path. Use the exercise to produce a short repair queue: missing identity mapping, unclear ownership, absent context, or an access path whose containment behavior still needs testing. Run it this week, before a vendor notice or a Wire item like this week's turns the exercise into the real thing.

Which of those gaps would your SOC hit first if it had to investigate one of your production agents today? Hit reply and tell me. I read every one.

See ya next week.

❝

Be good to each other, be kind to each other, love each other

David Lee

P.S. I made the quarterfinals of iHeartRadio's Top Host contest. The winner gets a six-episode podcast produced by iHeartRadio, $25,000 and a guest spot on On Air with Ryan Seacrest. Votes reset this round and only one host in my group moves on to the semifinals, so this is the round where your vote actually matters. It's free, you can vote once a day, and voting closes Thanksgiving night, November 26 at 10pm Eastern. Cast your free vote here. If the show has ever given you something you could use, that's what the ten seconds is for.

Reply

Avatar

or to participate