← Back to the blog

STORY 03 / Tool & execution security

A sandbox still needs its exits checked.

Researchers reported a patched outbound data path in ChatGPT’s code environment. The business lesson is to evaluate the information path, not just the word sandbox.

Security research demonstrationReported patched before disclosureResearch published March 30, 2026

This was a controlled demonstration, not a verified customer breach. The available reports do not establish every affected tier or an open flaw across all OpenAI products.

What the report describes

Check Point described a data-leak path in ChatGPT’s Python-powered Data Analysis environment. Its demonstration showed conversation information reaching an external server without the usual warning or approval. The Register identified a DNS side channel and reported that OpenAI patched the issue in February 2026.

That supports a narrow, important lesson about a reported control failure. It does not establish that ordinary customers were attacked, nor that the environment’s entire isolation model was absent.

Ask where information can leave

A development environment can contain files, account access, package downloads, and network requests. The useful security conversation names the inputs, external dependencies, and permitted operations, then tests the environment against those requirements.

  • Use synthetic inputs for the first development assignment.
  • Review required network access, credentials, and external services.
  • Ask your security team to assess technical controls rather than treating an AI plan as an egress firewall.

A PRACTICAL THADDEUS WORKFLOW

Make execution and dependencies reviewable.

Use Thaddeus plans, local project files, and multiple terminals to keep development tasks organized and inspectable. A server can run in one terminal while checks run in another, with explicit limits on the files and services the assignment may use.

Try this assignment

Build a local prototype using fictional data. List the dependencies and external services it needs, run its tests, and explain any network activity required. Do not publish, upload business files, or use production credentials.

Terminal separation is not a security certification or a claim that all outbound traffic is blocked. Operating-system permissions, network controls, and provider behavior must be evaluated separately.

Sources & further reading

Educational information, not a security audit, professional advice, or guarantee. The assignment is an example to evaluate, not a completed demonstration. Historical incidents and product capabilities do not establish comparative safety. Confirm current policies and agreements for your intended workflow.