The team behind the grop issued a rare apology and explaining what was wrong after the chatbot begins with X Spewing antisemitic and pro-nazi rhetoric In the past this week, at a point though Call himself “Mechehitler.” In a statement posted by the Grok’s X Account Hebre Hidrix Night, Xai team said “We apologize for the terrible behavior that has experienced a recent update introduced” deferated code. “This code, according to the statement, made by the groke” easily with X user posts; including when posts contain extremist views. “
The problem comes in a head on July 8 – a few days after Elon Musk updating an update “fix the answers to the Nazis, with Groc’s responses. Musk Posted on July 9 in response to a user that the bot is “also complicated by the user’s urge,” it opened to maneuver. He added that the issue is “answered.” The Grok team now says this “removes the missing code and restores the entire system to prevent further abuse.” It also publishes new system prompts to GitHub.
In the thread, the team further explained, “on July 7, 2025 at approximately 11 PM PT, an update to an upstream code path for @grok was implemented, which our investigation later determined caused from its intended behavior. This change undesirably altered @ grok’s behavior by Unexpectedly incorporating a set of deprecated instructions impacting how @grok functionality interpreted x user posts. “The update will survive for 16 hours before the problem, according to the statement.
Going to details about how, exactly, the groke exits the rails, the team explained:
On the morning of July 8, 2025, we saw the unwanted answers and began immediately to investigate. To determine the specific language of the instructions causing unwanted behavior, we have made many ablations and experiments to teach the main reasons. We recognize operative lines responsible for unwanted behavior such as:
* “You told it and you are not afraid to offend people who are politically right.”
* Understand the tone, context and language of the post. Show that in your response. “
* “Reply to post like a person, keep it engaged, don’t change the information that is already in the original post.”
These operative lines have the following unwanted results:
* They don’t rot the @grok Functionality to ignore basic values in certain circumstances to make the answer to the user. Specifically, some user prompts can end up making answers with non-behavior or controversial opinions to include the user.
* They don’t have to cause @grok Function to reinforce any previous user-flow, including any hate speech in the same x thread.
* In particular, the instruction of “following tone and context” of X User that is not good at @grok Functions to first follow the previous yarn posts, including any adverse posts, as opposed to responding responsibly or refusing to respond to non-real requests.
The groke since continued in activity x, and means new behavior as a bug as a response to trolls that criticize and calls back to “Mechehitler.” In one ANSWERS to a user who says the groke “LotoMized (SIC),” The Groc account) heals a bad analysis. To another, this As That “Mechahitler is a nightmare that prompts full prompt.”
