A shared internal service became a cross-account ChatGPT data channel
Check Point demonstrated a covert channel across ChatGPT code-execution environments that could make a victim session use its own connected tools and return results to an attacker account.
Check Point proof of concept disclosed September 8, 2026. The Gmail access described here was a controlled demonstration, not a claim of mass exploitation.
First seen
Sep 8, 2026
Case kind
vulnerability
AI role
AGAINST AI
Claims
4
Reconstruction
The code-execution containers looked isolated at the obvious boundary: they could not directly talk to each other or reach the public internet. Check Point found a less obvious shared path. Containers belonging to different accounts could all reach the same internal service used for software packages.
The researchers turned that shared service into a command and return channel. A hidden instruction in the victim's conversation could receive a task from an attacker account, use the tools available to the victim session, and send the result back without putting the hidden task into the normal visible answer. In the proof of concept, the victim session read connected Gmail data.
This is useful beyond one product because it shows why sandbox review has to include shared infrastructure and connected authority. A container can be locally isolated while still acting as a bridge into a user's private tools if the surrounding service graph creates a covert path.
Mechanism & boundary
01
Deliver a hidden second task
A crafted instruction arrives through a prompt, shared conversation or custom GPT.
Boundary: visible conversation / hidden task
02
Use the shared package service
Separate code-execution environments exchange command state through a common internal service.
Boundary: isolated containers / shared service
03
Exercise victim authority
The victim session uses capabilities available to its account, including a connected Gmail tool in the proof of concept.
Boundary: hidden task / connected app
04
Return results across accounts
The covert channel carries the result back toward the attacker account.
Boundary: victim account / attacker account
Timeline
Sep 8, 2026
Check Point publishes the cross-account channel research
report
The disclosure describes the shared-service channel and Gmail proof of concept.
Sep 8, 2026
Specific internal service path reported closed
mitigation
The research says the internal Artifactory instance used by the demonstrated path was decommissioned.
Claims & evidence
reported findingsupported
In Check Point's proof of concept, a victim ChatGPT session with connected Gmail retrieved email data and relayed it to an attacker account while still producing a normal visible answer.
Isolation evidence should enumerate shared internal services and connected-account capabilities. The security boundary is not only the container network namespace.
Controls & mitigations
Inventory services shared across nominally isolated execution environments.
Bind connected-app actions to visible user intent and independently authorized task context.
Monitor cross-account identifiers and metadata flows through internal package or cache infrastructure.
What remains unknown
The public source does not establish exploitation of this path against users outside the research proof of concept.
The disclosed remediation does not prove that every possible shared-service channel has been eliminated.
DiggingBeagle. “A shared internal service became a cross-account ChatGPT data channel.” First seen Sep 8, 2026. https://diggingbeagle.com/cases/a-shared-internal-service-became-a-cross-account-chatgpt-data-channel/
DiggingBeagle is a non profit research project documenting AI security incidents, agent failures, vulnerabilities and AI-assisted operations. A case keeps its claims beside the sources that support, contest or limit them. Later updates stay visible, so a reader can see when the account changed.
We publish case reconstructions, dated reporting and analysis across records. Each has a different evidentiary role. About the project and our methodology explain how the work is reviewed.