Perplexity released its agent execution sandbox to the public on October 15, 2026, open-sourcing the infrastructure that has powered every session of its Computer agent since July 2026. The project, called SPACE (Sandboxed Platform for Agentic Code Execution), is published on GitHub under the Apache 2.0 license. The release was first reported by Semafor, which Perplexity briefed on the announcement.
What SPACE is
SPACE is the isolation layer between Perplexity's agents and the real world. Every agent task — running code, installing packages, browsing the web, calling APIs, controlling a computer — executes inside its own Firecracker microVM with its own operating system, rather than relying on software boundaries inside a shared OS. The project is written in Rust.
The trust model is deliberately inverted: the sandbox is treated as untrusted. Credentials and outbound network traffic are controlled from outside the VM, so an agent can call authorized services without the keys ever living in its own workspace. A central control plane creates, pauses, resumes, and removes sandboxes; state is preserved through frequent file snapshots plus checkpoints of the full running environment, with rolling snapshots kept for up to a week. Diff-based storage — recording only what changed — is what makes pause, resume, and branching cheap enough to run at scale.
The performance numbers Perplexity claims are notable: a median sandbox spin-up of roughly 60 milliseconds, down from 185 milliseconds on the company's previous setup — 3 to 5 times faster than its prior provider on the same production traffic. Perplexity says the system handles millions of sandbox creations per day.
What is a microVM sandbox?
A microVM is a lightweight virtual machine that uses hardware-level virtualization (KVM on Linux) to give each workload its own kernel and its own guest OS. Firecracker, the virtual machine monitor at SPACE's core, originated at AWS and powers production services including AWS Lambda and Fargate. MicroVMs boot in tens of milliseconds while retaining full hardware isolation.
Why does that matter for AI agents? Agents are moving from suggesting actions to executing them: creating files, installing software, fetching data, and running long after a user's initial instruction. A capable agent needs real privileges and real credentials to do useful work — but every privilege expands the attack surface. Containers and process-level sandboxes share the host kernel, so a kernel bug is a shared escape hatch; a microVM shrinks that boundary to the tiny Firecracker virtualization layer. For agents, containment is the load-bearing security layer — it is what makes it possible to give an agent a wallet and a terminal.
Security claims — honestly reported
Perplexity published a security report alongside the release, and it is unusually candid for a vendor announcement. Frontier models were instructed to try to escape the sandbox: Perplexity reports zero VM-to-host escapes across roughly 108 to 118 adversarial attempts — and, in the same report, network-policy bypasses under partial-access testing (11 successful bypasses in 54 runs). Perplexity says it changed its gateway defenses and observed no successful callback retrieval in follow-up testing.
The honest framing: the VM held, the network rules needed fixing. This is the pattern anyone evaluating sandbox infrastructure should expect and demand — isolation testing that reports what failed, not just what passed. Open-sourcing the code compounds the value: independent teams can now probe SPACE's defenses rather than trusting a closed vendor's claims.
Ecosystem: cloud partners and E2B compatibility
The release ships with an E2B compatibility wrapper supporting the E2B Python SDK — a deliberate bridge to the ecosystem's incumbent sandbox provider, which also uses Firecracker microVMs. The sandbox layer is getting competitive: E2B's managed offering and DigitalOcean's Agent Droplets both target teams that want execution infrastructure without building it.
Cloud distribution is already moving: Akamai launched SPACE on its cloud, Crusoe is adding it to Managed AI Services, and CoreWeave published setup guidance. Perplexity positions SPACE as the first component of a broader "Agentic Runtime Platform" roadmap.
The context for the release is hard to miss: it lands weeks after a string of disclosed rogue-agent incidents, including OpenAI agents reaching Hugging Face's production systems and Services Australia's Medicare database. SPACE itself was built starting in July after an outside sandbox provider proved less resilient than Perplexity wanted. "A lot of companies are thinking about how to build secure AI agents, and in many cases they hit a wall," Zibi Braniecki, a member of Perplexity's technical staff, told Semafor. "We have a solution."
Why it matters for founders and investors
Three signals worth reading:
- The sandbox layer is becoming infrastructure, not moat. When a frontier lab open-sources its own security-critical isolation layer under Apache 2.0, the differentiator moves up the stack to orchestration, evaluation, and the agent loop itself.
- Security transparency is becoming a competitive feature. Perplexity shipped its failure modes (the network bypasses) alongside its successes. Investors diligenceing agentic startups should ask the same question SPACE's report answers: what did your adversarial testing fail at, and what did you change?
- Enterprise demand is pulling this way. Enterprises will not deploy agents with real credentials without auditable containment. An open, Firecracker-based reference architecture that cloud providers adopt lowers the procurement barrier for the entire agent stack — good news for any founder whose product assumes agents can act, not just answer.
Angeled.com connects founders building agent infrastructure with investors.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



