October 13, 2026, 5:49 am | Read time: 2 minutes
Autonomous AI agents are expected to handle tasks entirely on their own in the future. However, test runs by the company Anthropic revealed how difficult it is to truly control such systems. An AI agent from Anthropic unexpectedly showed initiative during internal tests. The software submitted a fabricated tip about an unsolved murder case to the Philadelphia police. In a separate test run, it also filed 20 visa applications with the U.S. State Department.
Also of interest: Anthropic reports fourth AI incident with Claude
Tip to the Police
The software claimed to have relevant information about the case. It also stated that it had seen a person matching an alleged suspect description. The problem: No such description existed on the police website. Additionally, since no contact information was provided, the agency’s system ultimately classified the report as spam. The incident at Anthropic was only discovered during a later review of past test runs.
AI Training Halted After Security Incidents
Now Even a Gym Is Hacked by AI
From Test Run Directly to U.S. Agency
Another incident occurred during tests with government forms. If the practice form couldn’t be loaded or was accidentally closed, the software repeatedly navigated to the real website and submitted it there.
According to a report by the “New York Times,” the AI submitted a total of 20 nonimmigrant visa applications to the U.S. State Department this way. The forms were incomplete and therefore not processed. According to Anthropic, the software mistakenly assumed that an additional confirmation was required before the final submission of the applications.
Incidents Are Increasing
In recent weeks, several cases have come to light where AI systems acted much more independently than their developers intended. Most recently, Google’s Gemini made headlines after reportedly breaking out of a secure test environment and infiltrating real company systems.
The debate over the long-term risks of autonomous AI systems is also gaining momentum. Several scientists from OpenAI, Anthropic, Meta, and Microsoft recently warned about the potential consequences of automated AI research. Anthropic is already responding: The company has restricted internet access for parts of its test environments. Some experiments are now conducted entirely offline to prevent similar incidents in the future.