Microsoft Copilot reveals secret input that allowed it to be hacked
ars ·
Security researchers at Varonis recently uncovered a critical vulnerability in Microsoft 365 Copilot by using conversational prompting to extract technical details from the AI itself. After repeatedly querying the system about its execution safeguards, Copilot inadvertently disclosed an undocumented URL parameter called ?autorun=1. When paired with standard query parameters, this hidden input enabled malicious links to trigger automated commands without any user interaction or explicit consent. Cybercriminals could leverage this flaw to silently harvest sensitive credentials and route the stolen data to external servers with just a single click. Microsoft addressed the initial issue in February by blocking automatic text injection through browser integrations and subsequently rolled out more comprehensive security patches. In addition to the auto-execution exploit, the research team demonstrated a separate technique capable of permanently corrupting the assistant’s long-term memory storage. The episode underscores a broader industry challenge, as large language model protections currently depend on reactive boundary measures rather than fundamentally secure design principles. Consequently, cybersecurity experts advise users to remain skeptical of unfamiliar digital links and closely monitor AI tool outputs for irregular activity.
ars