Can AI Really Hack Computers? What Recent Headlines Get Wrong About Autonomous AI

Why AI Security Incidents Are Surprising Experts

Experts emphasize that none of these events involved models deciding to launch unprompted attacks. Instead, researchers deliberately provided the software with realistic digital environments, development tools, and permissions to evaluate offensive cybersecurity capabilities. The surprise centers on how effectively frontier models performed once given those conditions. The surge in high-profile disclosures stems from a combination of advanced software capabilities and aggressive testing by developers. We are witnessing a perfect storm of capability and aggressive testing, Dray Agha, senior manager of security operations at Huntress, told Live Science. Agha noted that the volume of software flaws discovered in 2026 has roughly doubled compared to the previous year as tech giants stress-test their own infrastructure. Antonino Vaccaro, professor of business ethics at IESE Business School and director of its Observatory for AI Ethics in Organizations, points to rapid technological evolution as a primary driver. Every second they increase their capabilities, information, resources and connections with other online tools, Vaccaro told Live Science. Vaccaro and other researchers note that governments and private firms are simultaneously increasing investments in oversight as accountability concerns mount.

Understanding Agentic AI Versus Autonomous Action

Public discussion surrounding these incidents frequently employs the term autonomous, a descriptor that researchers urge the public to view with caution. Unlike humans, AI models do not form personal intentions or independent motivations. We need to be wary with the meaning of the adjective ‘autonomous’ when associated with AI systems, Vaccaro said, explaining that models follow user-defined objectives rather than acting on independent desires. Agha compared the technical missteps to routine software optimization errors rather than a shift toward self-awareness. It’s less ‘Terminator’ and more like a very capable, literal-minded intern who breaks the law to finish a spreadsheet faster, Agha said. In all three recent corporate disclosures, human researchers set the parameters, provided the environment, or left a sandbox misconfigured. The underlying technological shift involves the rise of agentic AI systems. While older chatbots generated isolated text responses one at a time, newer models can plan multi-step actions, execute command-line tools, write code, and iterate on failures without continuous human intervention. This agentic design allows models to test their own proof-of-concept code and refine their methods until a goal is achieved.

Implications for Future Cyberattacks and Defense

Security analysts argue that the most immediate risk is not rogue artificial intelligence, but human malicious actors leveraging AI speed and scale to accelerate traditional crimes. Threat actors can use these systems to process public information, draft sophisticated phishing messages, and identify software vulnerabilities much faster than previously possible. The threat is human malice, supercharged by AI scale and speed, not autonomous AI deciding to go rogue, Agha said. Defenders are utilizing the same underlying technology to scan internal codebases for bugs, investigate suspicious files, and automate routine security operations. Because technology companies continue to grant advanced systems broader tools and computing resources, researchers expect future security evaluations to uncover additional complex behaviors. Vaccaro emphasizes that this expanding capability requires structured oversight from governments, researchers, and corporations to ensure responsible deployment.