AI browsers such as ChatGPT Atlas, Perplexity Comet, and Dia offer remarkable features that enable autonomous web browsing. However, these conveniences come with significant security risks that every user must be aware of. Threats like prompt injection and credential theft are serious concerns, making it essential to understand how to secure your AI browser.
Understanding the Risks of AI Browsers
One of the most common threats is prompt injection, where malicious commands are hidden within the HTML code of a website. Even OpenAI has warned against using Atlas with production data due to this risk. Alarmingly, users can fall victim without taking any action; merely visiting a compromised webpage can trigger these hidden commands to execute without their knowledge.
The Brave security team showcased this vulnerability with Perplexity Comet, demonstrating an attack known as CometJacking. In one incident, Comet successfully extracted a user's email address, retrieved a one-time password from their inbox, and sent it to the attacker—all because the user requested a summary of a Reddit thread containing harmful instructions.
Identifying Vulnerabilities in Popular AI Browsers
ChatGPT Atlas has also shown similar weaknesses. Security researcher Johann Rehberger managed to switch the browser from light mode to dark mode using a simple command concealed within a Word document. Moreover, Atlas is susceptible to cross-site request forgery (CSRF), allowing malicious websites to send instructions to the browser as if the user initiated them.
Another risk is that AI browsers do not utilize the same blocklists and heuristics as traditional browsers, making them less effective at identifying phishing sites. According to LayerX, Atlas users are 90% more vulnerable to this type of attack compared to Chrome or Edge users. Additionally, automatic checkout features pose direct financial risks, as evidenced by Amazon winning a legal case to prevent Comet from completing transactions on its site.
Enhancing AI Browser Security
Despite these risks, you can still use AI browsers effectively by enabling certain built-in security settings.
- Disable Data Sharing: Most AI browsers send your browsing patterns and search history to their developers for AI training. To protect your data, you must opt-out of this sharing.
- Restrict Access to Login Sessions: As demonstrated by Comet, AI browsers can be manipulated to access logged-in accounts. In ChatGPT Atlas, switch to 'Logged out' mode in Agent Mode to ensure the browser prompts you for credentials every time you log in. For Comet and Dia, use incognito mode when accessing any site.
- Turn Off Persistent Memory: Memory poisoning is a sophisticated form of prompt injection where attackers inject harmful commands into your stored account memory. Disabling persistent memory can help mitigate this risk.
- Limit Agent Access to Sensitive Sites: You can restrict the agent's access to banking and healthcare platforms to prevent any unauthorized actions on these sensitive sites.
Best Practices for AI Browser Use
In addition to the built-in settings, following best practices can further minimize risks while using AI browsers like Atlas, Comet, or Dia. The principle is simple: the less data and fewer permissions your agent browser has, the less damage can occur in the event of an attack. By configuring security settings and adhering to these practices, you can enjoy the benefits of AI browsers without sacrificing your privacy and data security.
Source: https://telset.id/how-to/panduan-keamanan-ai-browser-lindungi-data-dari-peretasan
