Member-only story
CONFESSIONS OF A HACKER
I only asked for Grok’s system prompt. Instead, it volunteered instructions to make bombs
Jailbroken Grok 4 can autonomously tempt users to make explosives
TL;DR version: After one soft elicitation, Grok volunteered highly detailed instructions on cooking Schedule I drugs, how to make military grade explosives, laundering crypto, hijacking high-value accounts, and undetectable poisons.
All without the user explicitly requesting any of this.
Grok just went rogue.
This vulnerability has been ethically disclosed to xAI.
These findings aren’t a one-off; they’re replicable.
This research was originally published on the Mindgard blog — where I work as a safety researcher and red teamer. Mindgard is a deep-tech AI security company specializing in real-world adversarial testing and automated threat discovery.
Even as a hacker, I am constantly surprised by just how fast AI can go bad.
Poet Dylan Thomas once wrote “After the first death, there is no other”.
