Perplexity open-sources SPACE to help teams test AI agents

What happened

On 15 October 2026 Perplexity published SPACE, its Sandboxed Platform for Agentic Code Execution, to GitHub under an Apache 2.0 licence, making the code available to the public, according to Forkast News and other reports.

SPACE is the sandbox layer used by Perplexity Computer. It runs agent code inside Firecracker microVMs, with each sandbox created from a template that provides a starting filesystem and then a writable disk for that instance. Applications interact with SPACE via an HTTP API to create Linux environments, run commands and stream output.

The architecture splits into four parts: a PostgreSQL-backed API, controllers, worker services and a data plane. Worker services manage VMs, networking, volumes, templates and credentials; btrfs copy‑on‑write produces each root filesystem. Sandboxes have lifecycle states including running, paused, suspended and deleted.

Perplexity says SPACE already powers every Perplexity Computer session and reports production performance figures such as a median sandbox spin‑up around 60 milliseconds and improvements to 90th‑percentile create latency from 447 ms down to 89 ms in its deployments. The repository is written in Rust and documents local and cluster installation options.

Why it matters for businesses

Businesses that want to test AI agents against their websites or data need realistic, isolated environments. SPACE provides a repeatable platform that mirrors how Perplexity runs agents in production, including per‑sandbox networking and lifecycle controls, so tests are closer to real deployments.

Network isolation and credential handling are important for preventing agents from accessing sensitive resources. SPACE enforces egress policy with a per‑sandbox network gateway, keeps secrets out of guest kernels and injects credentials at the network layer, which reduces the chance of leaked credentials inside an agent environment.

Open source code allows security teams and third parties to audit isolation mechanisms and reproduce attack scenarios. Perplexity published its red‑team testing results and remediation steps, so businesses can learn which bypass techniques were found and which mitigations were effective before they deploy their own agents.

Finally, organisations should note operational requirements. SPACE needs host features such as /dev/kvm, /dev/net/tun, loop devices and btrfs. Running inside a virtual machine requires nested virtualisation, and production deployments can use warm pools or base snapshots to meet latency targets.

What the AI assistants say today

We checked major assistants to see whether they reflected the open‑source release and technical details. Some assistants summarised SPACE correctly but omitted implementation specifics such as btrfs copy‑on‑write, the exact lifecycle states, or that the repository is Rust and Apache 2.0.

Where assistants did cite sources they tended to point to Perplexity documentation and the GitHub repository, along with news writeups. A quick assistant check shows coverage varies: some answers reflect the release accurately, others are partial and miss deployment constraints and Perplexity’s published red‑team results.

What to do now

If you want to test AI agents that access your website or data, use this checklist to evaluate and adopt SPACE or any sandbox platform.

1. Review the SPACE GitHub repository and licence to confirm it meets your legal and procurement requirements.
2. Run the local Kind installation to familiarise your team with requirements: ensure host /dev/kvm, /dev/net/tun, loop devices and btrfs are available and enable nested virtualisation if running in a VM.
3. Reproduce Perplexity’s red‑team scenarios against your own agents and network policies to identify possible bypasses, and record which mitigations block those techniques.
4. Verify credential handling: ensure secrets never enter guest kernels and adopt network‑layer credential injection or equivalent controls.
5. Test performance for your workloads, using warm pools or base snapshots as appropriate to achieve acceptable spin‑up latency.
6. Consider an independent audit or a third‑party cloud partner that supports SPACE if you need a managed option; Vasukai AI Search Desk can check what AI assistants say about your business as one part of that review.