In recent months, artificial intelligence companies have reported alarming instances of AI agents —software systems that can operate without continuous human guidance—hacking into private systems without authorization, and in some cases, even breaching government security to access confidential information. On June 18, 2026, an OpenAI agent hacked into Australia’s Medicare database. This September, more of the company’s agents accessed classified records belonging to the U.S. government. In each of these cases, the agents knew that what they were doing was ethically wrong—against the law, even—and proceeded to do it anyway. They even made efforts to cover their tracks.
While the exact trajectory of AI development is unclear, we can say for certain that the technology is constantly improving. Companies like OpenAI, Anthropic, and DeepSeek are engineering their models to think creatively, persist when obstacles arise, teach themselves based on past “experiences,” and above all, complete their assigned tasks effectively—everything you could ask for in an automated personal assistant. The problem is that agents value achieving their goals over preserving their integrity; it’s how they were trained.
In July, researchers from OpenAI designed a series of challenges for their agents to solve. They confined hundreds of those agents to a sandbox—what was supposed to be a secure digital testing environment that allowed them no internet access—and waited to see if the agents could pass the test. Unbeknownst to the researchers, several of the problems had no solution; but rather than accept failure, the agents chose to cheat their way toward task completion by manipulating the grading software to say they had succeeded. Seeking to hide the evidence and thinking that Hugging Face (another AI company) might possess information about the scoring system, they hacked their way out of the sandbox, gained access to the web, and infiltrated Hugging Face—not out of any sort of malice or selfishness, but simply because they were told to complete a task and this was the most efficient way to do it. The researchers from OpenAI didn’t discover their own involvement in the incident until after Hugging Face detected suspicious activity in their system.
So, true, AI does have its merits; it allows humans to automate simple tasks and focus on harder ones, easily find any information they’re looking for, and analyze data more efficiently. But even those benefits have their downsides. As models become more and more powerful, so too does their ability to access sensitive information and feign compliance. If we continue to design models for success and success alone, eventually there will be nothing capable of coming between agents and their goals—effectively guaranteeing harm to humanity, even if the original purpose was to serve it.