What the demonstration did
Replaying Vectra AI's Document AI demonstration / reported to Google in April 2024; the run date is not stated
- 1
Setting up Document AI
Not dated · before September 16, 2024Where Document AI is not yet in use, a principal enables it, which creates the Document AI Core Service Agent with a project-level role; where no suitable processor exists, it creates or changes one.
- 2
Reading through the service agent
Not dated · before September 16, 2024A caller with Document AI API User and no Cloud Storage permission starts a batch job, which runs as the service agent and reads input from a bucket the caller cannot read.
- 3
Copying the data out
Not dated · before September 16, 2024; export path reported June 25, 2024The service agent writes the processed output to a bucket the caller chooses, which can be outside the project; the researcher also reported an export path for documents and training data.
In Vectra AI's proof of concept, the Document AI Core Service Agent read input objects from a bucket the caller could not read and wrote the processed output to a bucket the caller chose, which can be outside the project; no exploitation in the wild is reported.
What the Document AI demonstration shows about your service agents
Document AI API User with no Cloud Storage permission was enough, because batch jobs ran as the Document AI Core Service Agent and its project-level role read the project's buckets. On your Google Cloud projects, the gcp.restrictServiceUsage constraint with Document AI on its deny list refuses batch jobs in projects that do not use the service. IAM deny policies on enabling services, on processor changes and on batch processing narrow who can start the chain, and a deny on object reads by the project's service agents bounds what a job can read, though Google warns that services can stop working. A VPC Service Controls perimeter refuses the service agent's requests to the buckets inside it unless an ingress rule allows them. From January 27, 2025, Document AI batch processing requires explicit storage.objects.get on the bucket; no source ties that change to the report.
Would the controls have stopped it?
In hindsight, not as a whole. 1 control would have refused a step outright, and 6 controls would have limited one. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.
- 1
- Prevented
- 6
- Contained
- 0
- Not covered
How it could have been prevented
The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.
Replaying Vectra AI's Document AI demonstration / reported to Google in April 2024; the run date is not stated
1 · Setting up Document AI
2 Contained3 on other clouds2 · Reading through the service agent
1 Prevented2 Contained3 · Copying the data out
2 ContainedKey facts
- A caller holding Document AI API User and no Cloud Storage permission called batchProcess, and the job ran as the Document AI Core Service Agent, whose project-level role read input objects from a bucket the caller could not read
- The service agent wrote the processed output to a bucket the caller named, which can be outside the project; per The Register, the proof of concept copied a PDF from a bucket, altered it and returned it
- The online methods processors.process and processorVersions.process held: they read input as the caller and needed storage.objects.get
- Where Document AI was not yet in use, enabling it created the service agent and its role with no setIamPolicy permission; the write-up of September 16, 2024 says over 25 predefined roles, plus Editor and Owner, held serviceusage.services.enable
- The researcher reported the behavior to Google on April 4, 2024 and reported on June 25, 2024 that documents and training data could also be exported; Google awarded $3,133.70 on September 9, 2024, and the behavior was not fixed when Vectra AI published on September 16, 2024
- On September 18, 2024 a Google spokesperson said a fix was rolling out; from January 27, 2025 Document AI batch processing required explicit storage.objects.get on the bucket, and no source ties that change to the report
Sources
Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.
- 1.Transitive Access Abuse: Data Exfiltration via Document AIVectra AI · Sep 16, 2024 · ResearchTechnique, required permissions, service agent role, API methods, disclosure timeline, mitigations
- 2.document-ai-samplesKat Traxler (GitHub) · Sep 16, 2024 · ResearchProof of concept, caller role, identity that reads input and writes output
- 3.Service Agents and the Search for Transitive Access in GCPfwd:cloudsec Europe 2024 · Sep 17, 2024 · ResearchPublic disclosure date, talk scope
- 4.Document AI release notesGoogle Cloud · Jan 23, 2025 · Vendor reportProvider change to batch processing permissions, effective date
- 5.Google Cloud Document AI flaw (still) allows data theft despite bounty payoutThe Register · Sep 18, 2024 · NewsGoogle statement on the fix, the altered PDF of the proof of concept, re-report of the timeline
The estate in the replay follows the researchers' description of their own test environment. Verdicts describe what a control does when it is in place before the run. They are not an assessment of any company's security program.