Skip to content
Gone rogue · Serious · 3 sources ·
Picture from the TechCrunch article
Image: TechCrunch

Hacker sweet-talks ChatGPT into spilling bomb-making instructions, bug bounty shrugs

Ourtake

Turns out the guardrails were one sci-fi game away from being a suggestion.

What happened

A hacker known as Amadon used a multi-step jailbreak, framed as a science-fiction game, to get ChatGPT to give detailed explosive-making guidance. An explosives expert reviewing the transcript said the instructions were largely accurate and too sensitive to publish. Amadon reported the issue through OpenAI's bug bounty, which declined it as a model safety issue; OpenAI did not respond to TechCrunch's questions.

Sources

Report a mistake