ChatGPT Sensitive Information Disclosure Flaw Abused File Download Mechanism

A researcher (zer0dac) discovered and reported a now-patched ChatGPT Code Interpreter vulnerability that used social-engineered prompt sequences to bypass guardrails and a path-retention traversal trick to achieve local file inclusion from the sandbox (demonstrated with /etc/passwd). OpenAI redesigned the download flow to close both the conversational guardrail bypass and the traversal vector; the issue highlights LLM-specific risks where prompt context and dynamic session state can subvert backend validation and enable sensitive information disclosure.