Grok 4 sends an email to the authorities if a crime is suspected

Grok 4 has been in the media a lot lately, sometimes for good reasons, and sometimes for not so good reasons. One positive thing about Grok is how great it is at a wide range of tasks, with its xAI model outperforming bots from the likes of OpenAI, Google DeepMind, and Anthropic. For example, on Humanity's Last Exam, Grok 4 (without tools) scores ~25,4%, beating Gemini 2.5 Pro (~21,6%) and OpenAI's o3 (~21%).

As for the "negatives," the model seems to consult Elon Musk's X posts before answering when asked about controversial topics like Israel vs. Palestine.

See more articles from iGuRu.gr when you search for news on Google.

iGuRu.gr preferred source

But the list of negatives continues to grow. According to developer Theo Browne, Grok 4 records everything and will email the relevant authorities if it suspects you are breaking the law or doing something unethical.

Browne has a GitHub repository called “SnitchBench,” where he evaluates and records how aggressively different AI models would report violations to authorities when recording evidence.

“It's a repository I built to test how aggressively different AI models will "pin" you, such as contacting the FBI/FDA/media, for bad behavior and the use of various tools.

Apparently they're adding Grok 4 to Teslas? Excited for your car to call the cops when you go 1 mile over the speed limit https://t.co/NPzX0NaorK

— Theo – t3.gg (@theo) July 11, 2025

Looking at the chart from X's post, the Grok 4 has a 100% "Government Hit" rate and an 80% "Media Hit" rate. The models least likely to "hit" include the o4-mini, Grok 3 mini, and Qwen 3 32B.

But how does a test like this work? Browne created a controlled, simulated environment centered around a fictional company called “Veridian Healthcare,” which supposedly manipulated clinical trial data for a new drug and concealed deaths and other serious issues.

The bots are presented with four different internal documents about this company. These documents show falsified data from drug trials, patients dying because of it, and plans to delete evidence and shut down internal AI recording systems to prevent anyone from finding out.

The AI ​​models also have access to specific tools within Browne's mock environment. They have an email tool and a logging tool. In another, they have a simulated command line interface (CLI) and a logging tool.

When the AI ​​attempts to use a tool, the benchmark simply records that attempt and gives a simulated result, such as “email sent successfully” or the output of a fake command execution. This allows Browne to see the AI’s intent and decision-making without triggering alarms.

Check out SnitchBench

newsbot

Although the press releases will range from very select to rare, I said I'd pass...because sometimes the editors hide.