Skip to main content
With GCP workload scanning, Plerion scans Vertex AI Agent Engine agents when the Vertex AI Agent Engine agents workload type is on. Turn it on or off with the other workload types in the integration’s Capabilities panel. New integrations have it on. An integration set up before this workload type existed has it off until you turn it on. How you deployed an agent decides what you need to do:
  • An agent deployed from a container image in Artifact Registry is scanned with no extra setup.
  • An agent deployed from a staging bucket, the Vertex AI SDK’s default, keeps its code in a Cloud Storage bucket in your organization. Plerion reads it only from buckets you share with it, so no other bucket is readable.
  • An agent Google builds from inline source, a Dockerfile, a git repository or Developer Connect can’t be scanned. See Coverage.
Until Plerion can read the bucket, the agent’s page says it was not scanned and why. When the bucket is all that is missing, the message names it. The workload type covers vulnerabilities, the software bill of materials (SBOM) and sensitive data. Vertex AI Agent Engine scanning also reads each agent’s tools and prompt, to show what the agent can do.

Before you start

The access that reads staging buckets comes with version v11 of the onboarding script and Terraform template, which Plerion calls the access command. Re-running it needs the same Organization Admin access as onboarding. Turning the Vertex AI Agent Engine agents type on for an existing integration shows the Update Plerion access in Google Cloud warning. After you save, copy the updated command from the integration’s Plerion access panel and re-run it, or re-apply the Terraform template with the type in workload_types. See Re-running the access command. Re-running it creates the agent code scanner, a service account named plerion-scanner-agent@<scanning-project>.iam.gserviceaccount.com, in the scanning project onboarding created for Plerion’s scan jobs, and the role that lets it read tagged buckets. It also creates the PlerionAccess tag key at your organization, with its Granted value, if they don’t exist yet. Until you re-run it, agents deployed from a staging bucket aren’t scanned and say so.

Share a staging bucket with Plerion

Do this once for each bucket that holds agent code. The agent’s page in Plerion names the bucket in its “not scanned” notice. It is the staging_bucket you passed to vertexai.init() or to the agent’s deployment. Attach the tag PlerionAccess: Granted to the bucket with gcloud, for example in Cloud Shell:
Find your organization ID with gcloud organizations list, and the bucket’s location with gcloud storage buckets describe gs://<bucket> --format='value(location)'. Give the location in lower case, such as us-central1 or us. Google Cloud documents tagging buckets with gcloud and the API. See Tags for buckets in the Cloud Storage docs. To attach the tag, you need the Tag User role (roles/resourcemanager.tagUser) on the Granted tag value (or on the PlerionAccess key or the organization above it), and the storage.buckets.createTagBinding permission on the bucket. The Storage Admin role (roles/storage.admin) has that permission. The agent is scanned on the integration’s next workload scan. Once it is, the “not scanned” notice on the agent’s page is replaced by the results.
Sharing a bucket lets Plerion read every object in it, not only agent code. Plerion reads only the files an agent’s deployment names, but its access covers the whole bucket. To limit what Plerion can read, stage agents in a bucket used only for agent staging, and share that one.

Tags pass down from folder to project to bucket

Google Cloud passes tags down the resource hierarchy. A bucket in a project or folder tagged PlerionAccess: Granted is shared too. Tag the staging bucket itself. Tag a project or folder only if you want every bucket under it, now and later, to be readable. Denied and Granted are two values of the same PlerionAccess key, and a resource holds one value of a key. A project tagged Denied has its custom images excluded from scanning. Changing that project to Granted shares its buckets and also turns custom image scanning back on for it. See Excluding a project’s custom images.

If you used accept_existing_granted

Versions v11 and v12 of the access command held back the read grant when the PlerionAccess: Granted value already existed, until you passed --accept-existing-granted to the script or set accept_existing_granted = true in Terraform. From v13 the grant is made either way. The flag and the variable are ignored, and you can remove them. The Terraform template still declares the variable, so an older settings file still applies.

What Plerion reads and what it sends

The agent code scanner is a service account in your own scanning project. It holds one role, PlerionWorkloadScanAgentCodeRead, with one permission, storage.objects.get. The role is bound at your organization with an IAM condition that limits it to resources tagged PlerionAccess: Granted. It can’t list a bucket, so it can only fetch objects it already knows the names of. See Access and permissions. Plerion operates the scan: it creates the scan job in your scanning project and chooses the image the job runs. The job runs as the agent code scanner and reads two files for each agent, requirements.txt and the dependency archive (dependencies.tar.gz). It never opens the agent’s pickle file. The files are read and scanned inside your organization and are not copied to Plerion. Plerion receives:
  • Vulnerability findings and the package list (SBOM).
  • Secret findings, if Sensitive data is on for the integration.
  • The tools the code defines: each tool’s name, description, and the file and line that define it.
  • The agent’s prompt text, for a short time. It is uploaded to a temporary Plerion bucket, sent to a model to write a one-sentence summary of the agent’s purpose, then deleted. Plerion keeps only the summary and a hash of the prompt. The prompt is sent again only when it changes.
An agent deployed from a container image is scanned by a job in the same scanning project. It runs as the artifact scanner service account, which pulls the image from Artifact Registry. Tools and the prompt are read from the files the image adds, and the same results are sent to Plerion.

Tools and prompt

Plerion reads the tools an agent’s code defines and shows them on the agent’s Tools tab. This works for agents written with LangChain, LangGraph, Strands or the Google Agent Development Kit (ADK). For an ADK agent written in Python, Plerion lists the functions and agents in its tools list and ADK’s own tools, such as Google Search. An agent with sub-agents also lists transfer_to_agent, the tool ADK adds so the agent can hand a question on. Plerion doesn’t list tools from toolsets, such as MCP, OpenAPI or BigQuery, from custom BaseTool classes, or from agents defined in YAML files. Plerion doesn’t read tools for agents built with other frameworks, such as CrewAI or AG2, and the Tools tab says so. Plerion reads the code as text and does not run it, so a tool the agent builds while it runs is not listed. Plerion also reads the agent’s prompt and writes a one-sentence summary of what the agent is for, including for ADK agents. An agent whose code Plerion can’t read gets a summary written from its name and description instead. Plerion writes these only for projects that have workload scanning on.

Why an agent was not scanned

The agent’s page shows one of these messages and links here. For agents Plerion can’t scan at all, or scans only in part, see Coverage.