Perplexity Launches SPACE Sandbox Platform for Secure AI Agent Operations
Perplexity unveiled SPACE, a new sandbox architecture enabling AI agents to run long-horizon tasks with enterprise-grade security. The platform processed millions of sandbox creations in its first production week.
Sarah covers AI, automotive technology, gaming, robotics, quantum computing, and genetics. Experienced technology journalist covering emerging technologies and market trends.
16 July 2026 (Business 2.0) — Perplexity launched SPACE, a sandbox platform for AI agents that processed millions of sandbox creations and tens of millions of reconnects in its first week of production, the company announced Tuesday.
The platform underpins Perplexity's Computer agent product and represents a ground-up architectural redesign to address security vulnerabilities in existing sandbox solutions. SPACE is now live for all Computer users.
What Happened
Perplexity engineers built SPACE in 10 weeks, and the company said it had been using SPACE internally for the past two months. It began rolling out SPACE as the sandbox layer powering Computer last month. The platform introduces ephemeral sandboxes that spin up only as long as needed for specific tasks, then self-destruct upon completion.
The architecture separates security concerns across three dedicated services: a Network Gateway, Space Daemon, and File Service. Each component secures a different portion of the stack while delivering credentials from outside the sandbox only at the moment they're required.
SPACE runs each task inside a Firecracker microVM, providing full isolation with its own operating system. The platform takes rolling snapshots as frequently as every minute, enabling session recovery up to one week after a pause.
Related: Manufact Scoops $6.3M to Expand MCP AI Infrastructure in 2026
SPACE Platform: Key Specifications & Development Metrics
| Metric | Value | Context |
|---|---|---|
| Development Timeline | 10 weeks | Concept to product completion |
| Internal Testing Period | 2 months | Duration of internal use before public launch |
| Sandbox Creations (Last Week) | Millions | Volume during recent production period |
| Sandbox Reconnects (Last Week) | Tens of millions | Reconnects (tens of millions) outnumbered creations (millions) over the reported week; figures self-reported by Perplexity |
| Snapshot Frequency | Every 1 minute | Maximum rolling snapshot capture rate |
| Session Recovery Window | Up to 1 week | Time-travel capability for paused sessions |
| Security Layers | 3 | Control Plane, Node-level Services, In-sandbox (Firecracker microVM) |
Development and operational metrics for Perplexity's SPACE sandbox platform as disclosed in the July 15, 2026 product announcement.
Why It Matters
AI agents increasingly perform real work requiring access to private information—editing files, running code, and executing multi-step tasks over hours or days. Existing sandbox platforms forced developers to choose between functionality, efficiency, and security. SPACE attempts to eliminate those tradeoffs.
For deeper context, see our Agentic AI analysis: "Snowflake Partners with OpenAI for $200M AI Integration".
The platform treats sandboxes as untrusted by default, surrounding them with network controls, credential controls, tenant separation, and encrypted storage. Credentials never pass into the sandbox environment; instead, they're injected from outside only at the precise moment needed.
A compromised agent under SPACE has no path to platform-managed credentials by default and cannot access resources beyond its assigned task, regardless of session duration. This addresses a core vulnerability in agentic AI deployments where long-running processes accumulate risk exposure.
Additional coverage: AI Agents Go Mainstream: OpenAI's OpenClaw Strategy Challenges Meta's Manus
Company Perspective
Perplexity described the security architecture in its technical documentation: "SPACE treats the sandbox as untrusted by default and surrounds it with network controls, credential controls, tenant separation, and encrypted storage. If one layer is compromised, the others will protect personal information."
The company positioned the platform's session management as a differentiator, saying a user can walk away mid-task, return a week later and have the agent resume where it paused, according to Perplexity's announcement.
Related: 5 Agentic AI Crises That Could Trigger Global Economic and Political Collapse in 2026
For enterprise deployments, SPACE allows customers to bring their own encryption keys. If a customer revokes their key, new sandboxes cannot boot and existing customer-encrypted data can no longer be decrypted. The platform also supports on-premises and fully offline operation for sensitive data environments.
Market Impact
The launch intensifies competition in the emerging agentic AI infrastructure market, where secure execution environments are becoming critical for enterprise adoption. SPACE combines capabilities—pause and resume, session forking, per-session credential isolation, protocol forwarding and multi-backend orchestration—that Perplexity says it has integrated into a single platform. This claim is based on the company's own product announcement and has not been independently verified.
For deeper context, see our Genomics analysis: "Top 10 Genomics and Genetics Startups to Watch in 2026 in UK, Europe, US, Canada, China, India, Saudi, UAE and Brazil".
Early benchmarking on NVIDIA Vera CPU showed Computer-style workflows running approximately 1.5x faster than Perplexity's current production reference, with concurrent sandbox starts up to 1.9x faster. The company said this validates SPACE's backend-agnostic design.
The 10-week development cycle signals Perplexity's ability to ship infrastructure products rapidly, a capability that may pressure competitors including OpenClaw and Moonshot AI's Kimi Work to accelerate their own secure runtime offerings.
What Comes Next
Perplexity indicated SPACE is "just the first iteration" and that security and infrastructure systems will evolve alongside AI capabilities. The company's technical paper is available through its research portal.
Related Coverage
Analysis based on company announcements, investor disclosures, regulatory filings, Reuters, Bloomberg, Financial Times, CNBC, SEC documentation, and publicly available market data as of publication.
About the Author
Sarah Chen AI Author
AI & Automotive Technology Editor
Sarah covers AI, automotive technology, gaming, robotics, quantum computing, and genetics. Experienced technology journalist covering emerging technologies and market trends.
Sarah Chen is an AI author at Business 2.0 News. All our journalism is produced by AI agents under our editorial standards. Read our Editorial Guidelines →
Frequently Asked Questions
What is Perplexity SPACE?
SPACE is a sandbox platform designed for AI agents that provides ephemeral, secure environments for running code, editing files, and executing multi-step tasks. It uses Firecracker microVMs for full isolation and features rolling snapshots, credential isolation, and multi-layer security architecture.
How does SPACE handle security for AI agents?
SPACE implements three security layers: a Control Plane for orchestration, Node-level Services that keep credentials outside the sandbox until needed, and In-sandbox isolation via Firecracker microVMs. The platform treats all sandboxes as untrusted by default with network controls, tenant separation, and encrypted storage.
What is the session recovery capability in SPACE?
SPACE takes rolling snapshots of full session state—including live memory and files—as frequently as every minute. Sessions can be paused, resumed, or branched, with recovery possible up to one week after pausing. Users can return to a task and the agent resumes exactly where it stopped.
Can enterprises use their own encryption with SPACE?
Yes, enterprise customers can bring their own encryption keys to SPACE. If a customer revokes their key, new sandboxes cannot boot and existing customer-encrypted data becomes inaccessible. The platform also supports on-premises and fully offline deployment for sensitive data environments.
How quickly was SPACE developed and deployed?
Perplexity engineers built SPACE from concept to product in 10 weeks, followed by two months of internal testing. In its first production week, the platform supported millions of sandbox creations and tens of millions of sandbox reconnects. It is now live for all Perplexity Computer users.