AI startup Perplexity has raised alarms about emerging cybersecurity threats as artificial intelligence (AI) agents, which mimic human tasks, become more prevalent. These agents, capable of executing simple commands via plain language, may be manipulated by hackers using “query injection” techniques. Unlike traditional methods requiring technical expertise, modern attacks can originate from anyone proficient in language, leading to novel vulnerabilities. Experts recognize query injection as a significant threat, with notable figures like Eli Smadja calling it the “number one security problem” for AI-powered models. Major companies like Microsoft and OpenAI are investing billions into AI security, implementing defenses to detect malicious commands and requiring user supervision for sensitive tasks. However, the balance between security and user convenience remains challenging. Industry researchers warn that AI agents may not yet be reliable enough for critical functions, highlighting the urgent need for better safeguards in an evolving digital landscape.
Source link
