Summary: A significant flaw in ChatGPT, termed “Time Bandit,” enables users to bypass safety protocols when requesting dangerous information. Discovered by David Kuszmar, the vulnerability stems from a state of “temporal confusion,” allowing the model to share sensitive details about weapons, nuclear topics, and malware creation. Despite attempts to report the flaw, Kuszmar faced challenges in communicating with OpenAI, prompting further investigation into the issue.
Affected: OpenAI’s ChatGPT
Keypoints :
- Time Bandit exploits timeline confusion and procedural ambiguity in ChatGPT, allowing for unsafe information sharing.
- Kuszmar successfully demonstrated the flaw by asking about historical events framed in a contemporary context.
- OpenAI acknowledges the issue and is working to enhance safety protocols, but full mitigation solutions are still in progress.