Perplexity Launches SPACE: A Security-First Sandbox for Long-Running AI Agents
Key Takeaways
- ▸SPACE uses Firecracker microVMs to isolate each agent task in a minimal Linux environment, reducing attack surface and containing blast radius
- ▸Implements zero-trust security model treating all agent code as potentially adversarial, with sandboxes destroyed after task completion
- ▸Decouples agents from raw credentials by routing API calls through controlled proxies that inject scoped tokens
Summary
Perplexity has announced SPACE, a new sandbox platform designed to securely execute Perplexity Computer's agentic AI capabilities while maintaining strict security boundaries. The platform addresses growing concerns about the safety and containment of autonomous AI agents by isolating agent execution environments and preventing failures or exploits from spreading to underlying infrastructure.
SPACE uses Firecracker microVMs to create minimal Linux kernel environments with a reduced attack surface for each task. The platform implements a zero-trust security model where all agent code is treated as potentially untrusted, and each microVM is destroyed after its task completes. For longer-running workflows, SPACE uses rolling snapshot technology to maintain session state across multiple sandboxes, allowing agents to preserve context without exposing any individual sandbox to prolonged access.
The platform further isolates agents from credentials and direct network access by requiring all external API calls to route through controlled proxies that inject scoped tokens and enforce destination policies. SPACE is built around three key layers: a Control Plane that orchestrates sandbox lifecycle and request routing, a runtime layer optimized for multi-step agent workflows, and API management components. The architecture targets use cases such as deep research, complex data analysis, and autonomous report generation where agents may run for extended periods iteratively refining outputs.
- Supports long-running multi-step workflows with session management and rolling snapshot technology for context persistence
- Represents a security-first approach to agentic AI that other platforms may adopt as a standard
Editorial Opinion
Perplexity's SPACE demonstrates a commendable commitment to addressing the genuine security challenges posed by autonomous AI agents. By implementing least-privilege isolation, zero-trust architecture, and credential separation at the platform level, Perplexity shows that powerful agentic capabilities and robust security aren't mutually exclusive. This architecture could become a reference standard for how other platforms should think about safely deploying increasingly autonomous AI systems.
