Published: October 17, 2026 · Last updated: October 17, 2026
Perplexity open-sourced SPACE, the virtual machine sandbox it uses to contain its AI agents, saying its internal red team tried 118 different ways to break out of it and none worked.
Running an AI agent that can browse the web, write code, and take actions on a user’s behalf creates an obvious problem: what happens if the agent tries to do something it should not, either because it was tricked or because it simply went off the rails on its own. Perplexity’s answer to that problem, a sandbox environment it calls SPACE, is now open source, and the company is using a specific number to make its case for why that matters: zero successful escapes out of 118 attempts during internal red-teaming.
SPACE works by giving each agent session its own isolated virtual machine, with the ability to take snapshots of that environment so a session can be paused and resumed without losing its state. That sounds like plumbing, and mostly it is, but it is the specific kind of plumbing that determines whether a compromised or confused agent stays contained inside its own sandbox or finds a way to reach outside it, into the host system or a user’s actual files and credentials.
Perplexity said its internal red team attempted 118 distinct escape techniques against SPACE and none succeeded, a claim the company is now inviting outside researchers to try to break.
That last part, inviting outsiders to attack it, is the actual point of open-sourcing a security tool rather than keeping it proprietary. A company can run its own red team and report a clean record, and that claim means something, but it means more once independent researchers with no incentive to protect the company’s reputation get a chance to try the same thing. Security tools that survive public scrutiny tend to earn more trust than ones that only survive internal testing, and Perplexity is explicitly betting SPACE holds up under that kind of attention.
The decision also lands at a moment when AI agent safety has become a louder industry conversation, not a niche one. OpenAI, Anthropic, and Google have all shipped increasingly capable agents this year that can browse, click, and act with less human oversight than earlier chatbot interfaces required, and each added capability has come with a matching wave of security researchers probing for ways an agent can be manipulated into doing something its owner never intended. A shared, audited sandbox standard would be a genuinely useful thing for the industry to converge on, rather than every lab quietly building and testing its own in isolation.
Perplexity has separately signaled it plans to give away more of its underlying AI security tooling rather than keep it as a competitive moat, a notable choice for a company that is still competing hard against much larger rivals on the product side. The bet seems to be that being seen as the company taking agent safety seriously, in public and verifiably, is worth more right now than keeping the plumbing proprietary.
Related: OpenAI just shipped its own version of an agent that acts on your behalf inside the browser and how Google is handling a different kind of AI transparency question entirely.
Bottom Line: A sandbox that has not been tested by anyone outside the company that built it is a claim. A sandbox that is open source and has survived 118 attempted breakouts in public view is evidence. Perplexity just moved SPACE from the first category into the second, and that is worth more than the number itself.
Follow Teck Hustlers: Facebook · X · Instagram · LinkedIn · TikTok · Pinterest
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
