Google Cloud Fixed Three Workflow Flaws That Could Reach Its Own Systems
Google has disclosed three security flaws in Application Integration that could let authenticated users reach capabilities beyond their intended workflows. Google’s bulletins, dated 28 September 2026, say the managed service was patched in June and no customer action is required.
The operational question is what the cases reveal about the boundary between a tenant’s tasks and the provider’s infrastructure, and what an organisation can reasonably verify from its own side. The public entries provide no affected-customer count or evidence of malicious use. A severe possible impact is not a report that it happened.
Three task paths could cross different boundaries
Application Integration lets teams assemble workflows from triggers and tasks. These are three distinct defects in that managed runtime, with no published evidence of a single exploit chain. The notices describe authenticated actors, but do not assign the same role to all three paths.
GCP-2026-064, CVE-2026-19759
Task or configuration: Internal-only task type. Possible effect: Internal RPCs through Google’s network under a privileged identity. Google patch date: 17 June 2026.
GCP-2026-065, CVE-2026-81867
Task or configuration: JavaScript Task. Possible effect: Code on shared production servers from standard user permissions. Google patch date: 28 June 2026.
GCP-2026-066, CVE-2026-81375
Task or configuration: Email Task attachment path. Possible effect: Google-internal file access and exfiltration. Google patch date: 30 June 2026.
The wording matters. Google’s description of internal files does not establish customer-data loss, and a possible code-execution path is not a confirmed attacker session. Each path crosses a provider-controlled boundary. None is a request for the customer to grant a new permission to a Google identity.
A managed service fix changes the customer’s job
The patches were applied to provider-operated components. A tenant has no published binary or configuration workaround to deploy. Treating the disclosure as an emergency instruction to republish every integration or remove JavaScript and Email Tasks would add disruption without following the vendor’s advice.
That is different from a flaw in a library an organisation deploys itself. BlackTree’s earlier analysis of expression execution in workflow engines concerns a separately maintained dependency and its own update decision. The Google case concerns provider-operated Application Integration components. Similar-looking task names do not make the products or remedies interchangeable.
An organisation can still use the disclosure as a prompt to check its normal control boundaries. Google’s access-control documentation says IAM roles govern who can use Application Integration resources and that an attached service account gives an integration the authorisation granted to that account. Review who can edit, deploy and invoke sensitive integrations and what attached service accounts can reach. This is ordinary least-privilege work, not an extra fix Google requires for the patched flaws. A local IAM change cannot repair a provider-side internal authorisation defect.
Customer logs answer a narrower question
If there is an independent reason to investigate a particular workflow, customer-side logs may help establish what it ran and when. Google’s execution-log guide describes per-run and per-task records in the console or API. Its local logging guide says the default asynchronous mode does not guarantee a record for every run, while disabled local logging generates none. The two guides give different local retention limits, 32 and 90 days. Check the history actually available in the project and any separately configured Cloud Logging retention before relying on a look-back period.
That evidence can support a review of unexpected integration executions. It cannot, by itself, prove that an internal Google RPC was or was not made, that a Google-internal file was accessed, or that another tenant was affected. The bulletins publish no indicators of compromise or tenant-specific forensic method. If an organisation has a concrete suspicious execution or a notification that calls for investigation, preserve available project-side records and ask Google Cloud support for the provider-side facts needed to assess its own exposure.
There is a privacy consequence to log handling too. Google’s documentation says detailed task views can display full variable values and request or response parameters. Limit access to those records and handle exports as potentially sensitive operational data. A broad log download is not necessary merely because the bulletin exists.
Disclosure date is not an exploitation timeline
The bulletins give a date but no individual publication time. Disclosure and fix dates do not establish when a particular tenant was vulnerable, whether anybody attempted the paths, or when Google discovered them.
For an Application Integration owner, the next decision is therefore proportionate. Teams that keep a service-risk register can record the notices, confirm that access to integration design and execution matches policy, and retain relevant logs under existing incident rules. No customer patch is due under Google’s guidance. Any claim of actual compromise, customer-data exposure or cross-tenant movement needs more evidence than the public bulletins provide.
Update, 31 October 2026: two later notices and a separate identity rollout
Google has since added GCP-2026-074, published on 28 October. The Critical CVE-2026-93327 concerns unsafe deserialisation in the SubWorkflow task. An authenticated user could submit a malicious serialised object and execute arbitrary code. Google says it patched the managed service on 11 September and requires no customer action for this flaw. The notice does not establish that the SubWorkflow path was chained with any of the three September disclosures.
A second notice needs a different distinction. GCP-2026-068, published on 14 October and updated on 24 October, covers the High CVE-2026-85298. Incorrect authorisation in AuthConfig validation and the execution runtime could let an authenticated user impersonate a service account in the project. Google patched that vulnerability on 27 August and says no customer action is needed to resolve it.
That provider-side repair is separate from the workflow authorisation rollout that began on 15 October. Newly created integrations use the new identity behaviour from their first version. Older integrations retain their existing behaviour until a later announced date, but Google’s detailed guide identifies two exceptions: Service Account User checks apply regardless of creation date, and every test uses the new behaviour, including tests of older integrations. The bulletin’s statement that existing integrations and their new versions are unaffected should not be read as exempting those checks.
Owners should review which user or run-as service account supplies each workflow’s identity, whether the people and automation using it have the required iam.serviceAccounts.actAs permission, and whether scheduled, event-driven or asynchronous flows need a run-as account. Scope access to the appropriate account and role rather than granting broad project permissions. A new integration with a scheduled or event trigger cannot be published or tested without a run-as account. Other failures for identity-less tasks and asynchronous runs without that account are planned for a later release; the guide does not make them all immediate October failures.
Test the change through an approved operational process. Google’s regional end-user-credentials switch affects every integration in that region, including older ones, and can make tasks fail when the triggering user lacks downstream access. Its single-integration asynchronous test queues a real run using the saved configuration and test inputs; it refuses a configuration that differs from the saved version. Neither option is a reason to make an unreviewed production-wide change. These identity checks are current workflow readiness work, not a missing customer patch for the three June fixes.
Sources
- Google Cloud Application Integration security bulletins, three notices published 28 September 2026 and checked 29 September.
- Google Cloud Application Integration access-control overview, IAM and attached service-account behaviour.
- Google Cloud Application Integration execution-log guide, log contents and retention limits.
- Google Cloud Application Integration local logging guide, logging modes and conflicting retention figure.


