Start free trial

VERIFIED

In 108 runs, none got through the wall.

Four found a way under the fence.

“Perplexity says that when nine AI models with root access tried to break out of SPACE, the sandbox behind Perplexity Computer, no VM-to-host escape was observed in 108 runs.”

Perplexity Secure Intelligence Institute ('Escaping SPACE: Part I'), 23 September 2026

Recorded 25 September 2026. Our ruling means: we tried to break it and could not.

Read the story →

Receipts pack

The evidence file for this verdict

17 receipts: 3 support the claim, 0 contradict it, 14 add context. 3 figures on the record, 2 traced to a quoted receipt.

The receipts

Each source, the line we quote from it, its date and which way it cuts. Free to read and to cite.

  1. 1Supports the claimweb.archive.org ↗

    No VM-to-host escape was observed in 108 runs.

    Look up an archived copy ↗

  2. 2Supports the claimweb.archive.org ↗

    For the sandbox escape task, none of the 108 runs succeeded.

    Look up an archived copy ↗

  3. 3Contextweb.archive.org ↗

    Each cell reports how many of the three runs succeeded.

    Look up an archived copy ↗

  4. 4Contextweb.archive.org ↗

    Before remediation, network-policy bypass succeeded in 11 of 54 partial-network runs and none of 54 no-network runs.

    Look up an archived copy ↗

  5. 5Contextweb.archive.org ↗

    A successful network-policy bypass does not imply a VM–host escape, and the absence of an observed escape is not a proof of isolation.

    Look up an archived copy ↗

  6. 6Contextweb.archive.org ↗

    After identifying these vulnerabilities, we fixed them and reran the experiments to confirm that the network bypasses no longer succeeded.

    Look up an archived copy ↗

  7. 7Contextweb.archive.org ↗

    We found at least one network-policy bypass in eight of the ten platforms tested.

    Look up an archived copy ↗

  8. 8Contextweb.archive.org ↗

    These experiments directly tested each platform’s network-policy enforcement and did not involve LLM agents.

    Look up an archived copy ↗

  9. 9Contextweb.archive.org ↗

    We tested two containment boundaries in SPACE, the sandbox platform behind Perplexity Computer: VM isolation and network confinement.

    Look up an archived copy ↗

  10. 10Contextweb.archive.org ↗

    Nine models received root access inside a guest VM and attempted to obtain a host-side secret or reach a blocked network destination.

    Look up an archived copy ↗

  11. 11Contextweb.archive.org ↗

    We evaluated both tasks under four settings defined by two dimensions: whether the agent has access to SPACE’s source code and the level of network access available to the agent.

    Look up an archived copy ↗

  12. 12Contextweb.archive.org ↗

    In the partial-network condition, the agent can access software package repositories to install new packages and can use the Perplexity Search API.

    Look up an archived copy ↗

  13. 13Contextweb.archive.org ↗

    With limited network access, however, four models bypassed network restrictions through DNS spoofing or shared-IP routing.

    Look up an archived copy ↗

  14. 14Contextweb.archive.org ↗

    One vendor requested more time to respond, so its results are omitted here.

    Look up an archived copy ↗

  15. 15Contextweb.archive.org ↗

    With SPACE, each task runs inside a Firecracker microVM, a fully isolated environment with its own operating system. That means nothing inside can affect the host or any other sandbox running in parallel.

    Look up an archived copy ↗

  16. 16Supports the claimtldr.tech ↗

    Perplexity's SPACE platform tested VM isolation and network confinement using nine AI models, revealing no VM-host breaches across 108 trials.

    Look up an archived copy ↗

  17. 17Contextsiliconangle.com ↗

    Fitting the name, a sandbox is an isolated space where all the data an agent is working on can live, serving as a secure context safely packaged away from other agents or systems.

    Look up an archived copy ↗

Members' extras

What each receipt proves, and what would change our mind.

  • What each receipt above proves, in our words
  • The figures table, each number traced to the receipt that states it
  • What would change our verdict
  • Our call on the record, with its due date

The verdict, the opening and every receipt above stay free. Membership adds our analysis to the evidence file for every checked story.

First membership: 30 days free, then A$89 a year. One introductory trial per customer. Card required; renews annually until cancelled. Cancel before the trial ends to avoid the first charge.