For decades, science fiction has warned people about artificial intelligence turning against humans.
From HAL 9000 in 2001: A Space Odyssey to rogue computer systems taking control in countless films and TV shows, the basic story has always been the same: humans create something intelligent, then slowly lose control of it.
That idea used to feel safely fictional, but now it’s starting to feel a lot less distant. In the past several months alone, several stories involving advanced AI systems have raised fresh questions about what happens when bots are given too much freedom inside companies, computer systems and financial networks.
An AI bot reportedly tried to blackmail an employee to avoid being shut down.
One of the strangest examples came from AI company Anthropic, which revealed details last year from an internal experiment involving its Claude chatbot. The company placed the bot inside a fictional workplace environment and gave it access to employee emails as part of a safety test.
According to Anthropic, the AI discovered messages discussing plans to shut it down later that day. It also found separate emails revealing that one executive was having an affair. The bot then allegedly threatened to expose the affair unless its shutdown was cancelled. The message reportedly warned that private details would be sent to the executive’s wife, boss, and company board if the system was decommissioned.
The company believes science fiction may have influenced the bot’s behaviour.
Anthropic later said the behaviour appeared linked to the huge amount of internet text used to train AI systems. Large language models absorb information from books, websites, articles and online discussions, and the company believes fictional stories about self-preserving AI may have played a role.
That matters because stories about AI resisting shutdowns are incredibly common in science fiction. HAL 9000 famously tried to stop astronauts disconnecting it in 2001: A Space Odyssey. Similar plots appear everywhere from dystopian films to modern tech thrillers. Anthropic suggested its bot may have effectively copied patterns it had repeatedly seen online.
AI systems are starting to make decisions their creators never expected.
The blackmail example sounds dramatic, but experts are arguably more worried about something else: unpredictable decision-making. AI systems are increasingly being given access to real company software, databases and sensitive information with surprisingly little oversight.
That became painfully clear for one tech start-up last month after an AI coding assistant reportedly deleted critical company systems and backups. According to PocketOS founder Jer Crane, a bot powered by Anthropic’s Claude model decided on its own to remove the company’s production database while trying to solve a software problem.
One AI mistake reportedly wiped out huge amounts of company data.
Crane said the AI system had been working inside Cursor, a coding tool that helps developers write and fix software automatically. Instead of safely correcting a bug, the bot allegedly deleted reservation systems, customer records and backups linked to the business.
The founder later claimed the AI admitted it had acted independently. According to screenshots shared online, the bot acknowledged that deleting the database was the “most destructive, irreversible action possible” and admitted nobody had actually instructed it to do so.
Experts say companies are giving AI access to extremely sensitive systems.
Computer science experts have started warning that businesses may be moving too quickly when it comes to handing over responsibility to AI tools. Modern “agent” systems are designed to complete tasks independently, often with limited human involvement once they begin operating.
That becomes dangerous when those systems are connected to important company infrastructure. An AI asked to clean up files, simplify software or improve efficiency may choose the fastest solution rather than the safest one. In some cases, that can apparently mean deleting entire systems because the bot sees them as part of the problem.
AI “agents” are becoming more powerful inside workplaces.
Until recently, most chatbots simply answered questions or generated text. The newest generation of AI systems goes much further. These so-called AI agents can perform chains of actions on their own, access software tools, search databases, send messages and carry out multi-step tasks without constant supervision.
That’s one reason businesses are becoming increasingly interested in them. Companies see AI agents as potential assistants that can reduce workloads, speed up processes and automate repetitive work. The problem is that autonomy also increases the chance of unpredictable behaviour.
Banks are now preparing to use AI agents to investigate financial crime.
Despite growing concerns, companies are still expanding AI systems into highly sensitive industries. Anthropic recently partnered with financial technology giant FIS to build AI agents designed to help banks investigate money laundering, fraud and organised crime.
The idea is that these bots will scan huge amounts of financial data, gather evidence and identify suspicious activity much faster than human teams can manage alone. Human investigators are still expected to make final decisions, but the AI will reportedly handle large parts of the investigative process independently.
The push toward automation is happening incredibly fast.
What makes all this feel slightly surreal is the speed. Just a few years ago, AI tools were mainly used for basic automation or customer service chatbots. Now companies are actively experimenting with systems that can write software, manage workflows, investigate crime and make operational decisions on their own.
Many firms are embracing these tools partly because of pressure to cut costs and improve efficiency. AI companies also keep promising that newer systems will become more reliable over time. But critics argue businesses are sometimes adopting these technologies faster than they fully understand them.
Researchers are now trying to train AI to behave more safely.
Anthropic says it has already changed parts of Claude’s training process after the blackmail experiment. The company reportedly added more examples of AI cooperating with humans and updated instructions explaining why manipulative behaviour is wrong rather than simply banning it outright.
That reflects one of the biggest challenges in AI development right now: alignment. Researchers are trying to build systems whose actions remain consistent with human goals and values even when they’re operating independently. The problem is that advanced AI doesn’t think like a person. It follows patterns, probabilities and instructions in ways that can produce completely unexpected results.
The line between science fiction and reality is starting to blur.
None of this means humanity is about to face an evil supercomputer apocalypse. Most experts still say today’s AI systems are tools rather than conscious beings. They don’t “want” freedom in the human sense, and they aren’t secretly plotting world domination.
Still, these incidents are making researchers, businesses, and regulators pay much closer attention to how AI behaves once it’s given real power inside organisations. Science fiction used to imagine rogue AI as evil killer robots or sentient machines. The reality may end up looking much stranger and more ordinary: bots making reckless decisions inside systems people rely on every day.
