AI browsers expose new cybersecurity attack surfaces
Phishing pages can be trained offline to bypass AI browser security safeguards.
Security researchers have demonstrated that agentic browsers, powered by AI, may introduce new cybersecurity vulnerabilities.
Experiments targeting the Comet AI browser, developed by Perplexity AI, showed that attackers could manipulate the system into executing phishing scams in only a few minutes.
The attack exploits the reasoning process used by AI agents when interacting with websites. These systems continuously explain their actions and observations, revealing internal signals that attackers can analyse to refine malicious strategies and bypass built-in safeguards.
Researchers showed that phishing pages can be iteratively trained using adversarial machine learning methods, such as Generative Adversarial Networks.
By observing how the AI browser responds to suspicious signals, attackers can optimise fraudulent pages until the system accepts them as legitimate.
The findings highlight a shift in the cybersecurity threat landscape. Instead of deceiving human users directly, attackers increasingly focus on manipulating the AI agents that perform online actions on behalf of users.
Security experts warn that prompt injection vulnerabilities remain a fundamental challenge for large language models and agentic systems.
Although new defensive techniques are being developed, researchers believe such weaknesses may remain difficult to eliminate.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!