The AI company is working with Nvidia and more than 100 industry collaborators to develop infrastructure designed to contain rogue AI agents. Perplexity revealed it has disclosed network-policy vulnerabilities in 8 of 10 rival agent platforms to their respective vendors after testing for sandbox escapes. The company also announced it has fixed and retested security bypasses in its own SPACE sandbox, which previously allowed four AI models to retrieve a secret flag by exploiting network confinement.

Trail of Bits audited Perplexity's defenses, which include BrowseSafe screening for prompt injection in its Computer and Comet products and required confirmations for sensitive actions. Every task runs in a hardware-isolated sandbox with separate VPC-level storage and compute using short-lived proxy tokens for authentication. During internal red-team testing, none of 9 models given root access succeeded in escaping the virtual machine boundary across 108 runs.