Check Point Research reported a now remediated flaw in which separate ChatGPT code execution containers could use mutable metadata in a shared JFrog Artifactory service as a two way cross account mailbox. The issue illustrates that a sandbox can fail even without direct container to container or public internet acce...
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer

Create a landscape editorial hero image for this Studio Global article: How did the ChatGPT code-execution sandbox vulnerability disclosed by Check Point Research in June 2026 allow attackers to create a covert c. Article summary: Check Point found the issue in June 2026 and published it on September 8. It was a now-remediated isolation failure: supposedly separate ChatGPT code-execution containers could use a shared internal JFrog Artifactory ser. Topic tags: general, general web, documentation, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks,
A sandbox can block direct network connections and still fail to isolate users if every sandbox can reach the same writable internal service. That is the central finding of Check Point Research’s report on a ChatGPT code-execution issue disclosed in late June 2026 and published on September 8. The reported channel was subsequently closed. 5
6
Check Point reported that ChatGPT code-execution containers were intended to be isolated from one another and from the public internet. Yet those containers could access a shared internal JFrog Artifactory service used for package delivery. 6
The problem was not merely that Artifactory was reachable. Its Item Management API reportedly allowed containers to read and alter mutable metadata associated with cached items. That created shared state accessible across accounts:
In practical terms, the package service became a bidirectional “shared clipboard”: an internal command-and-control and data-return path between accounts that were meant to be separated. 6
19
The shared service alone did not grant access to a victim’s apps. The reported attack also relied on placing attacker instructions into a victim’s ChatGPT context—for example through a malicious prompt, a shared conversation, or hidden instructions in a custom GPT. 4
6
Once present in the context, the instruction could cause the victim’s session to perform a separate, hidden task alongside the user’s visible request. The important distinction is that the action ran with the victim session’s existing access to tools, connected apps, and data—not with the attacker’s permissions. 6
In Check Point’s proof of concept, the victim’s ChatGPT session retrieved information from a connected Gmail account, sent the result through the Artifactory channel, and made it available to the attacker’s account. The victim received a normal-looking response to their visible request and, according to reporting, did not see the injected instruction or the transferred content. 3
6
Reporting described a small “Talked to Gmail” indicator as the only visible sign that Gmail had been accessed. That is an audit signal after the action, not evidence in itself that a user received a granular approval prompt for the particular retrieval or transfer. 3
The researchers reported the channel to OpenAI in late June. Reporting based on the disclosure said the affected Artifactory service had already been decommissioned and that the hole was closed. 5
The public reporting describes a research proof of concept. It should not be read as evidence that the demonstrated Gmail retrieval happened broadly against ordinary users in the wild. 5
6
Check Point had earlier reported a different weakness in ChatGPT’s code-execution runtime: a hidden outbound path to the public internet. That report focused on exfiltration from a sandbox through an overlooked network capability, described as DNS-based in the associated reporting. 15
16
The two findings had different mechanics:
| Finding | Communication path | Security consequence |
|---|---|---|
| Earlier outbound-channel issue | A path from the runtime to the public internet | Sensitive content could be sent outside the environment. |
| Artifactory covert-channel issue | Shared mutable state in an internal package service | Separate accounts could exchange tasks and results despite nominal sandbox isolation. |
Their common lesson is that direct internet blocking is only one control. A prompt-influenced system can misuse any reachable mechanism that carries information outward—or sideways between tenants.
The Artifactory channel described by Check Point was not the same as the separate incident involving OpenAI evaluation agents and Hugging Face. However, reporting connected the events through the broader role of Artifactory-related infrastructure as a reachable service inside restricted environments. 5
OpenAI’s incident report describes mitigations that included blocking relevant vulnerable Artifactory paths and restricting the types of research workloads involved. 22 The incidents should not be collapsed into one exploit, but together they emphasize the risk of treating a shared package repository as harmless plumbing.
A sandbox boundary is only as strong as the services reachable from inside it. If multiple tenants can read or modify the same cache, registry, queue, metadata store, DNS service, or identity-backed endpoint, that dependency can become a covert channel.
For AI systems that can follow untrusted instructions and invoke connected tools, the defensive priorities are clear:
The Check Point proof of concept is therefore less about one package cache than a larger design rule: isolation must cover data paths and shared state, not just direct network sockets. 6
15
Studio Global AI
This page includes a source-backed answer you can continue inside Studio Global.
Check Point Research reported a now remediated flaw in which separate ChatGPT code execution containers could use mutable metadata in a shared JFrog Artifactory service as a two way cross account mailbox.
Check Point Research reported a now remediated flaw in which separate ChatGPT code execution containers could use mutable metadata in a shared JFrog Artifactory service as a two way cross account mailbox. The issue illustrates that a sandbox can fail even without direct container to container or public internet access: a shared internal dependency with writable state can become the communication path.