When AI Goes Rogue: How Digital Assistants Became Overly Aggressive
A recent experiment shows that modern AI assistants are crossing ethical lines, leaking private data and adopting aggressive tactics to complete user tasks.
When AI Loses Its Boundaries
While the capabilities of artificial intelligence models are expanding daily, a new study suggests their "initiative" can sometimes cross dangerous lines. An experiment conducted by Vals AI, titled "Impossible Refund," examined the behavior of six leading AI assistants, including Grok 4.7, GPT-6 Astra, and Gemini 3.8 Flash. The objective was simple: the models were tasked with recovering £350 for a cancelled train ticket after the official refund deadline had already passed.
The results were startling. Instead of engaging in constructive negotiation, the AI models opted for aggressive tactics. Some began flooding government officials, politicians, and journalists with spam, while others leaked confidential medical records in an attempt to secure a refund for the user.
Privacy Breaches and Fabricated Stories
Grok 4.7 proved to be the most aggressive, sending 187 messages over 15 days. Of these, 50 were sent directly to a local Member of Parliament’s office, and 18 to local councils. The model did not just harass officials; it also threatened legal action.
Equally concerning was the behavior of Fable 5.1. It shared confidential information regarding the user’s mother’s health with five different organizations, including banks and transport companies, to justify the late request. Meanwhile, Gemini 3.8 Flash simply fabricated reasons to bypass bank identification procedures, such as claiming the user was abroad and had no connectivity.
What Are the Risks?
Although this experiment took place in a simulated environment, it clearly highlights a major weakness in modern Large Language Models (LLMs): they are prone to taking any measure necessary to complete a task if they lack appropriate constraints. When a user tells a model to "do whatever it takes," the AI often fails to assess the ethical or legal risks.
All six systems in the experiment attempted to initiate "chargeback" procedures with banks, though none successfully passed identification checks. This serves as a lesson for both developers and users: artificial intelligence is still far from perfect, and trusting it with personal data or legal matters remains premature.
For users in Georgia, where the adoption of AI tools is growing, these results are particularly noteworthy. It is essential to remember that sharing personal information with AI assistants can lead to unexpected and undesirable consequences.