Autonomous software agents face fresh hacking attacks: hidden instructions can leak your data

Researchers who track software security say autonomous agents are running into a new wave of attacks. These digital assistants can read web pages, open files, summarize information, and carry out tasks, and that has sharpened concerns about how fast companies are rolling them out.

Those same researchers say agents that read web pages, open files, summarize information, and complete tasks can be manipulated by hidden or malicious instructions. The result can be data leaks, visits to dangerous links, or actions the user never meant to trigger.

They’re also warning about cross-app attacks. A poisoned calendar invite, spreadsheet, support ticket, web page, or document can be pulled from one service into another without proper checks, and the second app may treat it as harmless.

If you use autonomous agent features in your browser, phone, shopping app, or productivity service, take that seriously. The more access these systems have to your inbox, cloud storage, payment details, or work apps, the more damage a single bad instruction can cause. Be careful with broad permissions, especially when an agent can act on its own without a final confirmation.

Vendors still need stronger guardrails: clearer separation between trusted instructions and untrusted content, tighter default limits, better approval steps before money is spent, files are shared, or messages are sent, and more transparent logs that show why an agent took action. For now, these tools are spreading across consumer platforms as a mix of convenience and risk.