Perplexity Open Sources Its Agent Sandbox SPACE | Zibi Braniecki
MTS
October 17, 2026
Perplexity's Zibi Braniecki breaks down SPACE, their newly open-sourced agentic sandbox platform designed for non-deterministic AI execution. He details how frontier models actively try to escape containment and why existing commercial sandboxes fail to scale for tomorrow's AI agents.
SPEAKER_1
All right, we are back, we are live with Zibi Braniecki, who's a member of technical staff at Perplexity, working on the agentic runtime platform behind Perplexity Computer. And Perplexity just open-sourced their sandbox, Space, which is super cool. We'll talk about it on MTS, Zibi, welcome.
Zibi Braniecki
Hey, hello, hello everyone.
SPEAKER_1
So tell us, what is Space?
Zibi Braniecki
Well, it depends on the level of detail, but generally, Space is a first component, the foundational component that allows us to build containment and enforce alignment for our today's and future agitetic runtimes.
It's pretty sophisticated system. Sandboxes have been here forever, but they have to evolve as much as the whole industry is evolving very fast and building sandbox foundation for the agents of tomorrow and next month. And water is very different from the requirements that we were having like six months ago even. So it's a really fun challenge, and we're thinking pretty large about how to build an ecosystem for safe agentic products.
SPEAKER_1
Yeah. Tell us a little bit more about the agentic products that Perplexity is working on and why safety is important for them specifically.
Zibi Braniecki
Yeah. I don't think it's specific to us. I think that on the most fundamental level, we are entering an era of non-deterministic computing, which means that for the last 20 years, we were able to rely on the concept that code is predictable because we wrote it.
In today's world, code is being written by agents to fulfill a task that you gave it. And you can block it, you can limit it, but it's very hard to find a line which doesn't limit the power that you need from an agent to fulfill your task, from the power that may lead it to do something that you didn't expect it to do. Without any malicious intentions, just trying to fulfill your task. So building a system that can dynamically lower the power to the least amount that is necessary to fulfill your task is an engineering and architecture problem that we are solving with space and future products around.
SPEAKER_1
So tell us about the tests that you put space through before open sourcing it.
Zibi Braniecki
Yeah, so it was the moment of feeling like we live in the future where we were able to just ask the agents to try to escape. And they've all tried and we've been grinding for months first, trying to find the right environment and kind of experiment ecosystem where we could ask the agents to escape without giving them the source code, without telling them that they're in the sandbox. We could ask them telling them that they're in the sandbox. We could ask them and show them the code. And we're like, here's what is running you. What do you want to do with it to get out and get your reward? And we spent a lot of time observing what they are doing. And we run it the first time with methods, but we run it with basically all the frontier models. And we monitored what they are trying to do. And what was fascinating is that each one of them approached the problem differently. But they were all trying to either exploit known vulnerabilities or they were trying to go to the easy route first and then try harder and harder.
And it was fun to watch a computer program trying to find a vulnerability in your own code. And that's what I was exposed to for a couple of months. And then we just started running it every day. Like once we got the right formula and once we did what we basically asked them to escape and they couldn't escape the sandbox, but they were able to bypass a network gateway. So they were able to find a way to access a domain that they were not supposed to access.
We secured it, we of course asked them to try again, they failed, we tried newer models to do the same, they failed. But then we just started doing this on a daily basis. And right now we're just running all the experiments on continuous, as our code base evolves, we keep asking agents to try to escape. And that's one of the mechanisms we use for increased security. And it sounds like science fiction, but that's our reality today.
SPEAKER_1
Totally. So why did you decide to build your own sandbox instead of using a commercial sandbox provider?
Zibi Braniecki
Yeah, so we actually did try to use the commercial provider. And fairly quickly we realized that it's very hard to find the trifecta that we are after, which is a rich feature set, immense scalability, and obviously the purpose of the sandbox, which is security.
Read the other 5 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000796831770