Description

Perplexity AI launched Perplexity Computer, a cloud-based platform that enables users to develop and execute projects using advanced AI tools like Claude Code. Because the platform relies on generative AI services, it uses API tokens to allow the system to communicate with AI models and perform requests on behalf of users. A security researcher investigated the platform’s sandbox environment to determine whether these API tokens could be extracted. During the analysis researcher discovered endpoint addresses and active API tokens stored inside a .npmrc configuration file. By copying these tokens into external software researcher was able to make direct AI model requests outside the Perplexity environment. At first, these external requests appeared not to be reflected in the platform’s billing records, raising concerns that attackers could potentially exploit tokens to gain unlimited access to AI resources. The researcher shared the findings in a report titled “I hacked Perplexity Computer and got unlimited Claude Code,” which quickly spread across social media, particularly on X/Twitter. In response, Perplexity quickly revoked the exposed tokens and clarified that the tokens were not linked to shared corporate resources but were instead session-specific credentials generated individually for each user. Even when used outside the platform, any activity performed with those tokens would still be associated with the original user’s account and eventually billed to them. The reason the researcher initially observed no charges was due to asynchronous billing delays, which can take up to 18 hours to appear in the dashboard. According to Perplexity, revoking tokens was primarily a precaution to prevent researcher from accidentally accumulating high usage costs. The company also stated that showing tokens in environment variables is acceptable because they belong to user’s own session. Despite this explanation, the researcher argued that the design remains risky, warning that prompt injection attacks could potentially expose these tokens and allow malicious actors to misuse them, leading to unexpected financial liabilities for affected users.