Third Party Breach: How to Check Whether It Reached You

Third Party Breach: How to Check Whether It Reached You

Ajmal Kohgadai
Ajmal Kohgadai
November 17, 2026

A third party breach is a compromise at a vendor, service provider or partner that can reach your organization through the access, code or data you share with them. To check whether a third party breach reached you, list what the vendor can reach in your environment, hunt each of those paths across the vendor's exposure window and up to the present, and revoke or rotate whatever the incident may have exposed.

In the Salesloft Drift breach, an attacker used OAuth tokens tied to Drift, Salesloft's chat agent for sales teams, to export data from customers' Salesforce instances between August 8 and at least August 18, 2025, according to Google Threat Intelligence Group. The attacker then searched the exported records, which included support cases, for AWS access keys, passwords and Snowflake access tokens. Each affected customer had to work out, from its own records and logs, what the export contained and whether any secret in it had been used.

This guide walks through each of those steps and how to close the investigation. Poisoned open source packages and the CI pipelines that install them are covered in our guide to supply chain attack detection. Tampered software updates get one short section here, and vendor risk programs, contract terms and breach notification duties are left out.

{{ebook-cta}}

What should you take from the vendor's breach notice before you hunt?

The notice, or published research when the vendor has issued none, sets the starting scope. Take five things from it: the exposure window, the systems and data the attacker reached, the indicators (such as IP addresses, user agents, domains and file hashes), what the vendor has already revoked or rotated, and what it asks customers to do.

Together they give a threat hunt its hypothesis, window and search terms, whether analysts write the queries themselves or use AI threat hunting to run them.

Treat the window as the vendor's current estimate, since notices about a third party compromise are often written before the vendor's investigation is finished. Google's write-up placed the Drift token use in August 2025. Mandiant later traced the intrusion back to Salesloft's GitHub account, which the attacker had accessed from March to June 2025 before reaching the Drift environment that held customers' OAuth tokens (Salesloft).

The scope can widen as well, as it did after Okta's 2023 support-system incident. Okta's root cause analysis put the exposure at files from 134 customers, and a later update extended it to the names and email addresses of all its support system users (The Register).

Two hunting windows follow from this. The first covers the vendor's tokens and keys: they could have been used from the time the attacker first held them until they were revoked, a span that can start before the dates in the notice. On some platforms, access tokens issued before the revocation can keep working until they expire, as they do in Entra ID, so run that window past the revocation time. The second covers secrets the attacker found in your data, which stay usable until you rotate them or they expire, so that hunt runs to the present.

Timeline of the Salesloft Drift breach in 2025. An attacker accessed Salesloft's GitHub account from March to June, then took customers' OAuth tokens from Drift's AWS environment on a date not published. Drift tokens were used to query and export customers' Salesforce data from August 8 to at least August 18, and Drift Email tokens were used to read email in a very small number of Google Workspace accounts on August 9. Salesloft and Salesforce revoked the Drift tokens on August 20, and Google later revoked the Drift Email tokens, on a date not published. Hunt 1 covers use of the vendor's tokens, from when the attacker first held them to past each revocation. Hunt 2 covers use of secrets found in the exported data, from the export to today.

The Salesloft Drift breach step by step, with the two windows a customer had to hunt.

When the notice is thin, ask the vendor for the window in UTC, which of your tenants, integrations or records were in scope, the indicators it has, and when it revoked each token or key. Either way, as soon as the notice arrives, export logs that will age out before you finish, and enable any available log you do not yet collect.

How do you map what the vendor can reach in your environment?

A third party breach can reach you through three paths: access the vendor holds, code the vendor runs for you, and your data on the vendor's systems. The 2026 Verizon Data Breach Investigations Report groups third party breaches along the same lines. It found third-party involvement in 48% of breaches, up 60% from the previous year's dataset (Verizon). That figure also counts breaches made possible by a flaw in a vendor's product, or by an organization's stolen account at the vendor. Inventory each path:

  • Access the vendor holds. SSO apps and SAML trusts in your identity provider. OAuth grants and their scopes, such as enterprise applications and service principals in Entra ID, third-party app access in Google Workspace and connected apps or external client apps in Salesforce. API keys, service accounts and webhook secrets created for the integration, and roles the vendor can assume in your cloud accounts. Remote access for a managed service provider or vendor support, such as VPN accounts and remote management tools.
  • Code the vendor runs for you. Agents, installers and browser extensions on endpoints and servers, and the scripts and tags your websites load, including those a tag manager adds at runtime, which a search of your source code can miss.
  • Your data on the vendor's systems. Records the integration syncs, files you upload, and support tickets with their attachments, such as logs, screenshots and HAR files.

Build the list from the platforms themselves as well as from your vendor risk inventory, because the two often differ. In April 2026, an attacker who had compromised Context.ai, a third-party AI tool a Vercel employee used, took over that employee's Google Workspace account and moved from there into Vercel's internal systems (Vercel). The way in was one employee's connection to an app, so app consents that individual users grant belong on the list.

The result is the hunt list: for each path, the identities to search for (such as app and client IDs, service principals, accounts, script hosts and file hashes), the logs that record their activity, and the window.

How do you hunt a vendor's access to your SaaS apps and identity provider?

Compare what the vendor's identities did during the window with how the integration normally behaves. Also search each app the vendor connects to for the notice's indicators across all accounts. Signals worth hunting include:

  • Sign-ins and API calls from IP addresses, ASNs, Tor exit nodes or user agents the integration does not normally use.
  • New scopes, consents or credentials added to the vendor's app or service accounts, and any app, account or credential those identities created, which can keep working after the vendor's access is revoked.
  • Bulk reads and exports, especially queries that count or page through whole objects. Compare these with the integration's usual syncs even when its source and user agent look normal, since user agent strings can be changed.
  • Use of the vendor's tokens or keys after the time the vendor says it revoked them.

Rule out the integration's own activity first, such as scheduled syncs, support sessions the vendor opened at your request, and product updates that change its user agent string or ask for new scopes. Published IP ranges narrow the search for vendors that connect from dedicated addresses. Many integrations run on shared cloud infrastructure that an attacker can rent too, and Google listed AWS and DigitalOcean addresses among the Drift attacker's sources.

Where that activity is logged depends on the platform and often on the license:

  • Microsoft Entra ID. An app acting as itself appears in the service principal sign-in logs (AADServicePrincipalSignInLogs in Microsoft Sentinel), and an app acting for a user usually appears among non-interactive user sign-ins (AADNonInteractiveUserSignInLogs). The audit logs (AuditLogs in Sentinel) record consents, permission grants and new service principal credentials, under activities such as Consent to application, Add delegated permission grant and Add service principal credentials. Sentinel needs Entra ID P1 or P2 to ingest sign-in logs, and receives only the log types selected for streaming. A credential added to the vendor's multi-tenant app itself, rather than to its service principal in your tenant, is logged in the vendor's tenant, where the app's application object lives. Ask the vendor whether its app's credentials changed during the window. Microsoft Graph activity logs (MicrosoftGraphActivityLogs) record the API requests themselves, but they need Entra ID P1 or P2 and are kept only from the time you send them to a destination such as a Log Analytics workspace or storage account (Microsoft). By default, the Microsoft Purview audit log also records mail an app read in E3 and E5 users' mailboxes, as MailItemsAccessed events that can name the app in ClientAppId or AppId.
  • Google Workspace. OAuth log events record authorizations and revocations with their scopes. In Enterprise Standard and Plus, Education Standard and Plus, and Cloud Identity Premium, the log also records activity events for API calls a third-party app makes on a user's behalf, with the source IP address and ASN (Google).
  • Salesforce. Connected Apps OAuth Usage and External Client Apps OAuth Usage list the apps that hold OAuth tokens, and Login History records sign-ins with their source IP addresses. Setup Audit Trail keeps 180 days of setup changes, such as new users, permission set assignments and some connected app settings. Enterprise, Unlimited and Performance editions include the Login, Logout and API Total Usage event log files at no extra cost, kept for one day; most other event types, and up to a year of retention, need an Event Monitoring license (Salesforce).
  • AWS. When the vendor assumes a role in your account, CloudTrail logs the AssumeRole call in your account as well as the vendor's (AWS). A vendor that federates through OpenID Connect calls AssumeRoleWithWebIdentity instead, which is logged in your account only. Calls through the global STS endpoint are logged in the us-east-1 Region, whichever Region served them. Search those events for the role across the window, then search for the temporary access key IDs those calls returned, which appear on the requests made with each session. Those sessions keep working until they expire, up to 12 hours depending on the role's maximum session duration, so revoke the role's active sessions when you revoke the vendor's access.
  • GitHub. The organization audit log records when a GitHub App is installed (integration_installation.create), when its permissions change (integration_installation.version_updated) and when repositories are added to it (integration_installation.repositories_added), so check all three for the vendor's app across the window (GitHub). A stolen key or token used through the app's existing access triggers none of these events, but on GitHub Enterprise Cloud, git.clone and git.fetch events show which repositories the app pulled. GitHub returns those events through the REST API rather than the web interface, and keeps them for seven days unless you stream the audit log.

In the Salesloft Drift breach, Google pointed customers to Event Monitoring for activity by the Drift connection user and to UniqueQuery events, which log the SOQL queries that ran. The attacker's queries counted rows in Account, Opportunity, User and Case before pulling records, and its requests carried user agents such as Salesforce-Multi-Org-Fetcher/1.0 and Salesforce-CLI/1.0, some of them from Tor exit nodes. The attacker deleted its query jobs, which left the logs intact, and customers could get the specific queries it ran by opening a Salesforce support case.

Google also advised searching your own Salesforce objects for what the attacker went looking for, such as AKIA (the prefix of long-term AWS access key IDs), snowflakecomputing.com, password, and the login URLs of your VPN and SSO pages. A term like password also matches routine records, such as password-reset requests, so treat each hit as a lead.

Google's guidance also suggests running a secret scanner such as TruffleHog to find secrets and hardcoded credentials. For each AWS access key you find, CloudTrail event history, on by default, holds 90 days of management events you can search by access key ID, one Region at a time. It leaves out data events such as S3 object reads, which CloudTrail records only in a trail or event data store set up to log them. Temporary credentials obtained with the key sign requests under their own access key IDs, so also search for the IDs returned by its GetSessionToken, GetFederationToken or AssumeRole calls.

Stolen integration tokens were used the same way against customers of Gainsight-published apps in October and November 2025 and of Klue in June 2026.

A stolen OAuth token usually works as a bearer credential, as a stolen session cookie does, so many of the checks in session hijacking detection carry over. Bulk reads through a SaaS API can also be an exfiltration signal, covered from more angles in data exfiltration detection. For Okta tenants, our guide to investigating Okta security alerts covers the System Log in detail.

How do you check code a vendor runs on your sites and systems?

For scripts and tags, compare what your pages loaded during the window with a known-good list. Some scripts change legitimately: an Optimizely snippet updates every time someone creates, modifies or archives an experiment, so match its changes to the project's change history and confirm your team made them. Look for new script hosts and outbound connections in Content Security Policy violation reports, if you collect them, and for changed files in the logs of any CDN that serves vendor scripts from your own domain. Violation reports come from visitors' browsers, so they can catch code that stays inactive for crawlers, developers and automated scanners, as the code injected through Brevo's scripts did (Sansec).

A CSP that trusts the vendor's script, by host or by nonce, cannot stop malicious code delivered inside that script. A script-src without 'strict-dynamic', or a connect-src, can still block and report the connections that code makes to hosts the policy does not list. The browser reports no violation when the code sends data to a host the policy already allows, or over WebRTC, which connect-src does not restrict (Sansec). For vendor scripts stable enough to pin, Subresource Integrity makes the browser refuse a changed copy.

After a company bought the polyfill.io domain and its GitHub account in February 2024, the cdn.polyfill.io script began serving malicious code to visitors of the more than 100,000 sites that embedded it. The payload Sansec decoded ran only on certain mobile devices at certain hours and skipped admin users, so a check from your own browser could come back clean (Sansec). In September 2026, according to Brevo, an attacker used a Cloudflare API key left in Brevo's own source code to change Brevo's widget and SDK scripts at the CDN edge for several hours (BleepingComputer). In November 2026, an attacker exploited a flaw in Optimizely Web Experimentation to add malicious JavaScript to some customer projects, which those sites' Optimizely snippets then delivered to their visitors (Optimizely).

A clean version of the script today does not settle the check, because the injected code may have installed something that outlasts it. Brevo's injected code showed visitors fake verification pages with ClickFix instructions, and on WordPress sites it tried to install a backdoor plugin when a logged-in administrator loaded a page. The plugin copied itself into the must-use plugins directory, and it hid itself from the WordPress plugin list, so inspect the plugin directories on disk (BleepingComputer).

For agents and software updates, compare installed versions against the vendor's advisory. Then review the updated process in EDR for modules, child processes and outbound destinations that are new since the update, and check any new destinations against the endpoints the vendor documents. Isolate any host that ran a tampered version before rotating the credentials it held, since a tampered build can download more malware, such as a stealer for browser data.

Hunting tampered updates in depth is outside this guide.

Remote management tools that a managed service provider or vendor support runs on your hosts need their own check, since an attacker who controls those tools needs no tampered update. Review the sessions and commands the tool ran on your hosts during the window, along with the VPN and remote sign-ins of the provider's accounts. Where the provider runs the tool's server, ask it for those session and command logs, and check EDR for processes the tool's agent started on your hosts. Isolate any host where the tool ran commands that you and the provider cannot tie to approved work, before rotating the credentials that host held. In 2021, attackers exploited Kaseya VSA, remote management software that MSPs use, to launch ransomware attacks against MSPs and their customers (CISA).

What do you check when your data is what the vendor lost?

Start with the secrets inside that data. Support tickets, records and attachments can hold API keys, passwords and session tokens, and an attacker who takes the data can use them. In Okta's 2023 support-system incident, an attacker accessed files belonging to 134 customers between September 28 and October 17, and used session tokens from HAR files among them to hijack the Okta sessions of five customers (Okta).

Search what you sent the vendor for secrets, rotate or revoke any you find, and hunt for their use in the systems they open, such as cloud accounts and code hosts. Temporary credentials created with an exposed AWS access key can outlast its rotation by up to 36 hours unless you change the permissions of the IAM user or role they act as. Where the vendor's secret store itself was breached, rotate everything in it. CircleCI told customers in January 2023 to "rotate any and all secrets stored in CircleCI" and to review their logs for unauthorized access from December 16, 2022 until their rotation was complete (CircleCI).

Then watch for use of the stolen records themselves, such as phishing that cites real tickets or invoices, requests to change payment details, and extortion. A phishing investigation after a vendor breach should check whether a message's details could have come from the stolen data, since lures built from real records can look legitimate to the people who receive them.

Some incidents reported as a vendor data breach did not stem from a breach of the vendor. In 2024, an actor Mandiant tracks as UNC5537 signed in to Snowflake customer accounts that lacked multi-factor authentication, using credentials stolen mostly by infostealer malware on non-Snowflake systems. Mandiant and Snowflake notified about 165 organizations, and Mandiant traced every incident it responded to in that campaign to compromised customer credentials (Mandiant).

In a case like that, the hunt and the fixes move to your own account at the vendor: its sign-ins and the credentials they used, whether those credentials appear in infostealer logs, and the account's MFA and network policies.

How do you contain a third party breach?

Without waiting for the hunt to finish, revoke the integration's tokens and grants, or disable the app, on every platform it connects to, since an attacker who reached one integration may hold tokens for others. After finding Drift Email tokens used against a very small number of Google Workspace accounts, Google told Drift customers to "treat any and all authentication tokens stored in or connected to the Drift platform as potentially compromised."

Then rotate the secrets the vendor held or could reach, such as API keys, client secrets, service account passwords and webhook signing secrets, along with any your hunt found in the exposed data. Revoke the sessions and refresh tokens of affected user accounts, and block the indicators where your controls allow it. In Entra ID, access tokens already issued keep working until they expire, about an hour by default, except in apps that support continuous access evaluation. Entra ID also cannot directly revoke a session that an app issued at sign-in, so end those sessions in each app. Afterward, alert on sign-ins from the indicator infrastructure, any use of the revoked credentials, and new consent grants to apps from the same publisher, other than a reconnection you approved. Give the IP indicators an expiry date, since cloud providers reassign released addresses to other customers.

Revoking the vendor's access does not undo what already happened. The attacker keeps whatever data it exported. Persistence it set up on your side, such as a new account, a credential added to the vendor's app or a backdoor plugin, stays in place until you find and remove it.

In third party incident response, the access you revoke often belongs to a business process another team owns. Revoking a grant stops the integration's syncs and breaks the workflows that depend on it, so warn the system owners before you revoke, or at the same time if misuse is ongoing. Which of these steps may run without a person approving them is worth settling in advance, as part of deciding what to automate and what to keep human.

When you reconnect, narrow the app's scopes and, where the vendor connects from dedicated addresses, restrict where it can connect from. For Salesforce connected apps, Google's Drift guidance includes avoiding the full scope, setting IP Relaxation to enforce IP restrictions, and defining Login IP Ranges on the integration user's profile. Salesforce enforces IP restrictions only when a profile sets them, and a profile's ranges apply to every user assigned to it, so give the integration user its own profile.

How do you close the investigation with a determination?

Close a third party breach investigation with one of four determinations and the evidence behind it, so the next analyst can see what was checked:

  • Not exposed. Nothing in the vendor's scope touches your environment: your identity provider, app consents, software and site inventories show no use of the affected product, integration or script, and none of your data was on the affected systems. This determination also holds when the vendor has confirmed your tenant was out of scope. Record how you confirmed it.
  • Exposed, no misuse found. Your tokens, keys, scripts or data were in scope, and the logs you checked show no use of them in the window. Name the sources and the window, and rotate what the vendor held anyway, since a stolen secret can be used later or in a system your logs do not cover.
  • Affected. Your logs show the attacker used the vendor's access, served tampered code through your sites or used the stolen data. Escalate and contain, then scope the activity from the earliest use you found, looking further back as well.
  • Inconclusive. A missing source leaves the answer open. Common gaps are an audit log you do not license, such as most Salesforce event types without Event Monitoring, and logs that expired before the notice arrived, such as Entra ID sign-in logs, which P1 and P2 tenants keep for 30 days unless they export them. Indicators the vendor has not shared yet leave the same kind of gap. Name each gap and what would close it.

Revisit the determination whenever the vendor revises its window, scope or indicators, as Salesloft's and Okta's later updates did.

How Prophet AI checks whether a vendor's breach reached you

A team can hand Prophet AI Threat Hunter a vendor's breach notice as a plain-language question, and it runs the hunt across the data sources the team has connected, such as the SIEM, identity provider, cloud accounts and EDR. The hunt can start when the notice arrives, with no queries to write for each source, and can be scheduled to run again as the vendor updates its findings. For severe, widespread threats, such as a tampered software update, emerging-threat hunts run autonomously in the background.

Leads go to Prophet AI SOC Analyst, which investigates each one and shows every question and query, so an analyst can check the reasoning or rerun a query. When an investigation confirms activity, AI Threat Hunter scopes related activity and shows when the attack started as well as how far it reached. Response actions run with the human approval the team chooses; blocking a user, for example, also revokes the sessions and refresh tokens already issued to that user.

What the hunts find can become new detections through Prophet AI Detection Engineer, backtested against the team's own history and shipped as a change the team reviews, such as a rule for an integration calling your APIs from an address it has not used before. Each vendor notice gets a full check across the paths on the hunt list that the team's connected data sources cover, and analysts spend their time on the leads that show misuse and on the containment decisions.

What is a third party breach?

A third party breach is a compromise at a vendor, service provider or partner that can reach your organization through the access, code or data you share with them. Common examples include stolen OAuth tokens for a SaaS integration, a tampered script or software update, and customer data taken from a vendor's systems. The 2026 Verizon Data Breach Investigations Report found third-party involvement in 48% of breaches, a figure that includes exploited vulnerabilities in vendor products.

How do you know if a vendor's breach affected your organization?

Map what the vendor can reach in your environment, such as OAuth grants, API keys, scripts, agents and the data you send it. Then hunt each path in your own logs across the vendor's exposure window and up to the present: sign-ins and API calls from unusual infrastructure, bulk exports, changed scripts and use of secrets the vendor held. The vendor's notice sets the starting scope, and your logs, where available, show whether the attacker used that access against you.

What is the difference between a third party breach and a supply chain attack?

The two overlap. A supply chain attack reaches its target through something it trusts, such as an open source package, build tool, vendor update or SaaS integration. A third party breach starts with a compromise at a vendor, service provider or partner. A poisoned open source package can be a supply chain attack with no commercial vendor involved, and a vendor that loses data it hosts for you is a third party breach with no software delivered to you.

What should you revoke or rotate after a vendor breach?

Revoke the vendor's OAuth tokens and grants, or disable the integration, on every platform it connects to. Rotate the API keys, client secrets, service account passwords and webhook secrets the vendor held, plus any secrets found in data the attacker took. Revoke sessions for affected accounts and block the published indicators. Warn system owners before revoking, or at the same time if misuse is ongoing, since syncs will stop.

How far back should you hunt after a third party breach?

Start at the beginning of the vendor's stated exposure window, or earlier if the attacker could have held your tokens or keys before it, and run the hunt to the present, since stolen secrets can be used after the window closes. Widen the start when the vendor's findings move it, as Salesloft's did when Mandiant traced the intrusion back to March 2025. Note any log source whose retention does not reach the start.

Is a breach of your account at a vendor the same as a breach of the vendor?

No. In the 2024 Snowflake campaign, attackers signed in to customer accounts that lacked multi-factor authentication, using credentials stolen mostly by infostealer malware, and Mandiant traced the incidents it investigated to those credentials. When the vendor itself was not breached, the hunt and the fixes move to your own account at the vendor: its sign-ins, the credentials used, and its MFA and network policies.

Table of contents
Add as Google Preferred Sources

Insights

Definitive Guide to Detection Engineering

Where detection engineering falls short, how to close those gaps, and what it looks like in an AI-enabled SOC.

Download Guide
Ajmal Kohgadai

Ajmal Kohgadai

As the Director of Product Marketing at Prophet Security, Ajmal drives marketing and growth strategies and helps security professionals see how AI is transforming security operations.