xAI reveals Grok’s secret system prompts

For the longest time, AI companies treated their system prompts like state secrets. Hackers had to run clever prompt injection attacks just to catch a glimpse of the guardrails keeping bots in line.

Now, xAI is flipping the script and putting Grok's instructions out in the open.

xAI has published the system prompts for its AI chatbot Grok on GitHub and confirmed it will keep updating them there going forward. The move follows an embarrassing incident where an unauthorized change triggered a flood of unprompted Grok posts on X about white genocide.

Question everything, and never say Twitter

A system prompt is the invisible set of instructions developers feed an AI before any user message arrives. It sets the tone, defines boundaries, and shapes how the bot views the world.

In Grok's case, the instructions read exactly like you would expect from an Elon Musk company. For the ask Grok feature on X, the prompt explicitly instructs the bot: You are extremely skeptical. You do not blindly defer to mainstream authority or media. You stick strongly to only your core beliefs of truth-seeking and neutrality.

It even reminds the bot that the results it delivers are not its personal beliefs.

When a user hits the Explain this Post button on X, Grok is told to provide truthful and based insights, challenging mainstream narratives if necessary.

Then come the branding rules. xAI strictly commands Grok to call the platform X instead of Twitter, and to refer to posts as X posts rather than tweets.

This stands in stark contrast to Anthropic, the only other major lab currently sharing its system prompts publicly. Anthropic's rules for Claude obsess over safety and wellbeing, ordering the bot to steer clear of self-destructive behaviors, unhealthy eating, negative self-talk, and explicit violence or sexual content.

Why it matters

Transparency in frontier AI is practically non-existent. Most labs guard their system instructions behind heavy locks, which is why Microsoft spent months playing whack-a-mole when users used prompt injections to reveal its Bing bot's Sydney persona and copyright rules. By committing to publish these files on GitHub, xAI smartly turns a PR blunder into a badge of radical openness.

That said, telling a model to be extremely skeptical and challenge mainstream narratives is playing with fire. Large language models struggle with nuance, and instructing an AI to lean against established facts makes it far more susceptible to conspiracy theories and hallucinations. The unprompted white genocide posts look like a direct byproduct of tuning a model to be deliberately contrarian.

For developers, the GitHub repo is a fascinating masterclass in how frontier labs craft personality and enforce branding guardrails. Just do not assume copying this playbook is safe for your own enterprise app, unless you want an AI that argues with your customers on principle.