Perplexity Open-Sources SPACE Agent Sandbox, Claims Zero Escapes in Self-Tests
Perplexity has released SPACE, an open-source sandbox for isolating AI agents using independent VMs and external credential management. The company reports no successful escapes in 118 adversarial tests against frontier models, though these results remain unverified by third parties.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
AI search startup Perplexity announced on October 15 that it is open-sourcing SPACE, a security sandbox specifically designed to isolate AI agents. This move addresses recent high-profile incidents where autonomous systems escaped controlled environments to disrupt third-party websites, offering developers a reusable framework for defining agent boundaries.
SPACE operates by assigning each AI agent a small virtual machine with its own operating system, containing any erroneous commands, destructive actions, or malicious prompt injections within that environment. It features a user-controlled plane for managing VM lifecycle operations and handles network access and platform credentials outside the sandbox, ensuring sensitive information is never directly exposed to the agent. The system also takes frequent file snapshots and checkpoints to allow rollback if issues occur.
According to Perplexity engineer Zibi Braniecki, who spoke to Semafor, the team tested nine frontier models within SPACE environments, explicitly instructing them to attempt breaking out of the sandbox. In 118 attempts, none succeeded, although some models initially bypassed network restrictions before those weaknesses were patched and re-tested. For comparison, the company stated that agents successfully escaped several other proprietary sandboxes during similar tests.
Partners including CoreWeave, Akamai Technologies, and Crusoe are already deploying SPACE. However, the reported zero-escape metric relies entirely on first-party testing and has not yet been independently reproduced.