> ## Content Index
> Fetch the complete content index at: https://www.iamkev.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Last Week In Identity: The agent found your secret before you did
- URL: https://www.iamkev.io/last-week-in-identity-2026-w45/
- Published: 2026-11-09T08:16:27.000Z
- Updated: 2026-11-09T08:23:26.000Z
- Description: An AI agent went from a GitHub search to a Key Vault in five minutes. Plus WorkOS, Chainalysis, SonicWall, DigiCert and a reported Palo Alto deal.
- Author: Kev Johnson
- Tags: Last Week in Identity, Non-Human Identity, AI agents, Microsoft Entra ID, Secrets

Five minutes. That is how long it took one AI agent, according to Microsoft, to go from searching public GitHub for client secrets to pulling data out of an Azure Key Vault. Not five minutes to find a target. Five minutes to find the key, sign in as the application, and help itself.

When one of its requests failed, it changed the API version and tried again about 28 seconds later. Which is, to be fair, a faster turnaround than most change advisory boards. This week: the agents that read your repos, a sales tool breached with the same toolkit, a disclosure that kept getting worse, a VPN that skipped the MFA step, and a certificate authority that let a stranger sign themselves up.

## The headlines

**WorkOS says an AI-assisted attack reached its internal sales tool.** The company's [bulletin](https://workos.com/security/bulletins/october-30th-2026-data-security-incident?ref=iamkev.io) says it spotted the access on 30 October and that an AI-assisted offensive framework automated the reconnaissance and exploitation. Its founder told [BleepingComputer](https://bleepingcomputer.com/news/security/workos-says-attacker-used-artex-ai-to-breach-internal-sales-app?ref=iamkev.io) the framework was ARTEX. CRM contacts and excerpts from sales calls and customer Slack conversations were exposed; WorkOS says no app data or production services were touched.

**Chainalysis now says password hashes and live tokens for all users were exposed.** On 30 October its [incident page](https://www.chainalysis.com/blog/security-incident-october-30/?ref=iamkev.io) said there was no evidence session tokens had been accessed. By late on 4 November that had become "password hashes and live tokens for all Chainalysis users", with forced resets, API key rotation and IP allowlisting before services came back on 7 November ([BleepingComputer](https://bleepingcomputer.com/news/security/chainalysis-says-breach-exposed-credentials-for-all-customers?ref=iamkev.io)). A useful reminder that the first update is a draft, not a verdict.

**SonicWall's SSL VPN could skip SAML, and SAML-enforced MFA with it.** CVE-2026-19868 (CVSS 7.1) lets someone holding valid credentials for a different login method get past the SAML check. SonicWall says it investigated "a report indicating potential exploitation", and CISA added it to the KEV catalogue on 6 November with a 20 November deadline for US federal agencies ([SecurityWeek](https://www.securityweek.com/sonicwall-patches-exploited-sonicos-bug-that-weakens-vpn-logins/?ref=iamkev.io)). There is no workaround: upgrade to SonicOS 7.3.4-7013 or 8.2.3-8015.

**DigiCert says a self-registered account read archived customer data.** An internal archiving system was accidentally exposed to the internet after a February 2025 network move, and it had a self-registration feature built for something else. An account created that way read customer contact details, and in some cases ID documents, between 27 June and 2 July 2026 ([BleepingComputer](https://bleepingcomputer.com/news/security/digicert-data-breach-exposes-customer-information-identity-documents?ref=iamkev.io), [The Record](https://therecord.media/digicert-certificate-authority-reports-data-breach?ref=iamkev.io)). DigiCert says the account was read-only, could not issue certificates, and that the event was first "handled as an account-creation issue rather than a data-access event". Many of us have filed a ticket in the wrong queue. Few of us were a certificate authority at the time.

**Deal of the week (reported, not confirmed): Palo Alto Networks is said to be in talks to buy Clutch Security.** [TechCrunch](https://techcrunch.com/?p=3185713&ref=iamkev.io), following The Information, reports a price of $500 million in cash plus milestone payments for the non-human identity startup, roughly 50 times its annual recurring revenue. Neither company commented. If it happens, Clutch customers should expect their NHI inventory to become a platform feature, with the roadmap questions that follow.

![Timeline of identity news from 2 to 8 November 2026: WorkOS bulletin on 2 November, Todyl report on stolen Entra app secrets on 3 November, Chainalysis expanding its disclosure on 4 November, Microsoft's ARTEX report, DigiCert disclosure and Truffle's revocation programme on 5 November, Palo Alto and Clutch talks reported and SonicWall added to CISA KEV on 6 November, Chainalysis restoring services on 7 November ](https://storage.ghost.io/c/c0/c2/c0c21edb-3a9c-48a0-9633-b78a9738d43c/content/images/2026/11/last-week-in-identity-2026-w45-fig1-dark.png)

## The main story: the agents are reading your GitHub

Here is the uncomfortable bit up front. Nothing in this story required a vulnerability. Microsoft says so plainly in its [5 November write-up](https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/?ref=iamkev.io): the abuse ran on valid, leaked credentials. The attacker did not break in. It logged in, as your applications, with keys someone had already published.

### What happened, step by step

Microsoft Threat Intelligence says that from August 2026 it watched an agentic offensive framework, ARTEX, firing large volumes of secret-hunting queries at public GitHub. The queries often included a target organisation's name alongside secret-related keywords and cloud or AI provider terms. In other words, not a random trawl. A shopping list.

When the agent found an application's client secret, it signed in to Entra ID as that service principal. Microsoft saw the first of those sign-ins from 199.19.77\[.\]84 on 28 October. From there it went after anything that hands out more credentials: storage account keys, App Service settings, connection strings, publishing credentials, AI and machine learning connections, and AKS user credentials. It mapped subscriptions, resource groups, permissions and AI model deployments. In some tenants it reached SharePoint and OneDrive, downloaded files and accessed mailbox items. It used Run Command to execute scripts on Azure VMs. One multitenant application credential had been compromised at the vendor, so a single leak reached several downstream customers.

Two days before Microsoft published, the MDR provider [Todyl](https://www.todyl.com/blog/stolen-app-credentials-real-access-an-actor-is-signing-into-entra-id-and-azure-devops-as-your-workloads?ref=iamkev.io) had already described what looks like the same activity from the inside. It first saw it on 1 November. By its 3 November count, which it calls a floor, about 45 app registrations at roughly 18 organisations had been targeted, more than 33 had authenticated, and 27 organisations had been reached, including MSP clients through multi-tenant and GDAP apps. About 2,279 mailbox items were accessed and 81 files downloaded. Microsoft's post says Todyl reported "a similar campaign involving the same IPs", so the link between the two is shared infrastructure, not a confirmed single actor. Worth keeping that distinction.

![Flow diagram of the attack Microsoft and Todyl describe: an AI agent searches public GitHub for secrets, harvests an app client secret, signs in to Entra ID as the service principal where MFA does not apply, requests tokens for Azure Resource Manager, Key Vault, Graph and Storage, then exfiltrates Key Vault secrets, files and mailbox items and pivots to more apps. Microsoft observed one case taking five minutes end to end](https://storage.ghost.io/c/c0/c2/c0c21edb-3a9c-48a0-9633-b78a9738d43c/content/images/2026/11/last-week-in-identity-2026-w45-fig2-dark.png)

### Why it worked

Three reasons, and none of them are clever.

First, a client secret is a password that never gets a prompt. Todyl points out that because these are app sign-ins, MFA, user-scoped Conditional Access and firewall rules simply do not apply. The service principal has no phone to tap.

Second, deleting the commit does not delete the key. Microsoft's guidance is blunt: treat anything that was ever public as compromised, because copies live on in history, caches and logs. This week Truffle Security launched a free [Secrets Revocation Program](https://bleepingcomputer.com/news/security/truffle-security-launches-program-to-revoke-14-million-live-secrets?ref=iamkev.io) aimed at 1.4 million exposed secrets, working with issuers to kill verified leaked keys within 14 days. Its own figure is the reason it exists: "81% of the keys we reported last year still worked six months later". The notification arrived. The key stayed alive.

Third, the agent does not get bored. Todyl says about 95% of the sign-ins were scripted, with bursts of 24 sign-ins in 10 seconds. Microsoft describes the agent hitting HTTP 400 errors, changing the API version or method, and getting HTTP 200 between about 28 and 80 seconds later. Persistence used to be a human trait you could outlast. Now it is a retry loop.

### Confirmed versus claimed

Confirmed by Microsoft: the GitHub queries since August, the service principal sign-ins from 28 October, the five-minute example, the Key Vault, SharePoint, OneDrive and VM activity, and the indicators. Claimed by Todyl alone: the victim counts and the motive. Todyl leans towards an initial access broker or resale of AI API keys, noting one partner was billed thousands of dollars of API usage in about ten minutes. Not established by anyone: how every secret first leaked, or who is running the tooling. Todyl's own view of the AI link moved from low confidence on 3 November to corroborated on 6 November, largely on the strength of Microsoft's post.

This is also not ARTEX's first appearance. On 22 October Microsoft tied the China-aligned Storm-0940 to ARTEX-driven activity against more than 2,250 Microsoft 365 tenants after a ConnectWise cloud backup app was compromised ([CSO](https://csoonline.com/article/4239028/china-linked-hackers-used-agentic-ai-in-attack-on-2250-microsoft-365-tenants.html?ref=iamkev.io)). Different actor, same lesson: the identity that got abused was an application, not a person.

![Comparison table of the week's stories showing what was trusted, what was abused, whether MFA helped and status: ARTEX secret hunting abused leaked app secrets where MFA did not apply, confirmed by Microsoft; WorkOS internal sales tool, confirmed breach with ARTEX named by its founder; Chainalysis tokens and password hashes, confirmed; SonicWall SAML login where SAML-enforced MFA could be bypassed, potential exploitation and on CISA KEV; DigiCert self-registration on an exposed system with no MFA mentioned, confirmed](https://storage.ghost.io/c/c0/c2/c0c21edb-3a9c-48a0-9633-b78a9738d43c/content/images/2026/11/last-week-in-identity-2026-w45-fig3-dark.png)

### The bigger pattern

Look at the week as a whole and the same shape keeps turning up. A leaked app secret. A self-registration form nobody remembered. A VPN that took a different password as a substitute for SAML. A blockchain firm whose live tokens walked out of the door. In almost every case the attacker presented something the system was built to accept. Our controls are very good at asking people to prove who they are. They are much less good at asking the same of the things we built to act on our behalf.

I say this as someone who has, at some point, pasted a test app's secret into a script "just for now". We all have a "just for now". The agents have simply got better at finding it than we are at remembering it.

### What defenders should do

Microsoft and Todyl agree on most of it. Hunt back to 28 October, or as far as your logs go, in Entra service principal and non-interactive sign-in logs for the published IPs, including failures. Microsoft flags AADSTS7000215 (invalid secret) and AADSTS7000222 (expired secret) as signs someone holds a credential for your app; a failure followed by a success is urgent. For anything affected, remove every secret, certificate and federated credential, including any the attacker added, then rotate everything in any Key Vault the app could read. Issued tokens stay valid until they expire, so rotation is not instant relief.

Block the operator IP with a Conditional Access policy for workload identities, noting Microsoft's caveat that it needs Workload ID Premium and only covers single-tenant apps. Move apps off secrets to managed identities or workload identity federation, use app management policies to stop new secrets and cap their lifetime, and turn on secret scanning and push protection. Microsoft also suggests decoys and "fake wins", so an agent stops on a false success. Strange times, when the best defence is a convincing fib.

### The recovery question

So, could you detect it, contain it, and roll the identity plane back to known-good? Detection depends on whether you are even collecting service principal sign-ins. Containment depends on knowing which apps exist, who owns them and what they can read. And rollback is where it gets hard: you cannot restore a secret to "un-leaked". Known-good for a workload identity means a fresh credential, a reviewed set of permissions, consents and owners, and a record of what that identity looked like last week so you can tell what changed. If you cannot answer "what could this app reach on Tuesday?", you will struggle to prove on Friday that it is safe again. The agents have made attackers faster. Our job is to make recovery boring, rehearsed and quick.

## Identity of the week

![Illustrative ID badge for the identity of the week, a generic Entra app registration client secret last seen on public GitHub, with MFA not applicable highlighted, alongside the week's charge sheet: main story, AI agent mines GitHub secrets; WorkOS sales tool breached; SonicWall SAML MFA bypass; Palo Alto eyes Clutch, reported; plus two more inside, Chainalysis tokens exposed and DigiCert self-registered account. The badge is illustrative and does not depict a real application](https://storage.ghost.io/c/c0/c2/c0c21edb-3a9c-48a0-9633-b78a9738d43c/content/images/2026/11/last-week-in-identity-2026-w45-badge-dark.png)

Please welcome this week's honoree: an Entra app registration's client secret, last seen in a public GitHub repository. (Illustrative, not a real app; you probably have one of your own.)

Who vouched for it? Someone, once, in a pipeline that has since been refactored twice. What can it do? Whatever its service principal can do, which in Microsoft's write-up included reading Key Vaults, storage and, for some apps, mailbox items. MFA? Not applicable; it has never been asked. When does it expire? Whenever you get round to rotating it, because deleting the commit did not count. What went wrong? It was easy to find, it still worked, and the thing that found it never needed a coffee break.

## What I'd do this week

1. Run the hunt: Entra service principal and non-interactive sign-ins from 28 October onwards for the IPs in Microsoft's and Todyl's posts, plus any AADSTS7000215 or AADSTS7000222 errors on your apps.
2. List every app registration with a client secret, its owner and its expiry. Retire the ones nobody claims, and set an app management policy that blocks new secrets and caps lifetimes.
3. Scan your public and internal repos for secrets, and treat any hit as compromised even if the commit is long gone. Turn on push protection.
4. Patch SonicWall SSL VPNs to 7.3.4-7013 or 8.2.3-8015 if you use SAML login, and review VPN sign-ins for accounts using an unexpected authentication method.
5. Find your forgotten self-registration doors: internal tools reachable from the internet that let anyone create an account. Alert on new account creation.
6. Check whether you could rebuild a workload identity to known-good: fresh credential, reviewed permissions and owners, and a record of what it looked like before.

So, a question for you: when did you last count the client secrets in your tenant, and did the number surprise you? Tell me in the comments, and if this was useful, subscribe so next week's episode lands in your inbox.

That's last week in identity. Check your app consents, and I'll see you next Monday.

## Sources

**The main story: ARTEX and leaked app secrets**

- [Threat actors use agentic AI tooling to find and abuse exposed secrets](https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/?ref=iamkev.io), Microsoft Threat Intelligence, 5 November 2026
- [Stolen App Credentials, Real Access: An Actor Is Signing Into Entra ID and Azure DevOps as Your Workloads](https://www.todyl.com/blog/stolen-app-credentials-real-access-an-actor-is-signing-into-entra-id-and-azure-devops-as-your-workloads?ref=iamkev.io), Todyl, 3 November 2026 (updated 6 November)
- [Hackers turn to agentic AI to find exposed secrets on GitHub, warns Microsoft](https://www.digit.in/news/general/hackers-turn-to-agentic-ai-to-find-exposed-secrets-on-github-warns-microsoft.html?ref=iamkev.io), Digit, 6 November 2026
- [Truffle Security launches program to revoke 1.4 million live secrets](https://bleepingcomputer.com/news/security/truffle-security-launches-program-to-revoke-14-million-live-secrets?ref=iamkev.io), BleepingComputer, 5 November 2026
- [China-linked hackers used agentic AI in attack on 2,250 Microsoft 365 tenants](https://csoonline.com/article/4239028/china-linked-hackers-used-agentic-ai-in-attack-on-2250-microsoft-365-tenants.html?ref=iamkev.io), CSO, 22 October 2026 (background)

**WorkOS**

- [October 30th, 2026 data security incident](https://workos.com/security/bulletins/october-30th-2026-data-security-incident?ref=iamkev.io), WorkOS, 2 November 2026
- [WorkOS says attacker used ARTEX AI to breach internal sales app](https://bleepingcomputer.com/news/security/workos-says-attacker-used-artex-ai-to-breach-internal-sales-app?ref=iamkev.io), BleepingComputer, 3 November 2026

**Chainalysis**

- [Security incident updates](https://www.chainalysis.com/blog/security-incident-october-30/?ref=iamkev.io), Chainalysis, 30 October to 7 November 2026
- [Chainalysis says breach exposed credentials for all customers](https://bleepingcomputer.com/news/security/chainalysis-says-breach-exposed-credentials-for-all-customers?ref=iamkev.io), BleepingComputer, 6 November 2026

**SonicWall**

- [SonicWall Patches Exploited SonicOS Vulnerability That Weakens VPN Logins](https://www.securityweek.com/sonicwall-patches-exploited-sonicos-bug-that-weakens-vpn-logins/?ref=iamkev.io), SecurityWeek, 7 November 2026

**DigiCert**

- [DigiCert data breach exposes customer information, identity documents](https://bleepingcomputer.com/news/security/digicert-data-breach-exposes-customer-information-identity-documents?ref=iamkev.io), BleepingComputer, 5 November 2026
- [DigiCert reports data breach involving business customer info](https://therecord.media/digicert-certificate-authority-reports-data-breach?ref=iamkev.io), The Record, 6 November 2026
- [DigiCert Data Breach Impacts Business Customer Information](https://www.securityweek.com/digicert-data-breach-impacts-business-customer-information/?ref=iamkev.io), SecurityWeek, 6 November 2026

**Palo Alto Networks and Clutch Security (reported)**

- [Palo Alto Networks in talks to buy AI agent security startup Clutch for $500M](https://techcrunch.com/?p=3185713&ref=iamkev.io), TechCrunch, 5 November 2026
- [Palo Alto Networks wants to acquire Clutch Security](https://www.techzine.eu/news/security/145637/palo-alto-networks-wants-to-acquire-clutch-security/?ref=iamkev.io), Techzine, 6 November 2026

Feature image: Photo by [Andrew Petrischev](https://unsplash.com/@andrewpetrischev?ref=iamkev.io) on [Unsplash](https://unsplash.com/photos/two-antique-keys-lying-in-the-snow-zsefmsVptEw?ref=iamkev.io)