Every good prison-break movie needs a warden who swears the place is escape-proof. Perplexity just played that role in public, handed the blueprints to everyone, and invited nine of the smartest inmates on the planet to give it their best shot.

118 Escape Attempts, Zero Jailbreaks

Perplexity has open-sourced SPACE, the sandbox software it uses to contain AI agents, so outside developers can now use it too, Semafor reports. Cloud and edge providers CoreWeave, Akamai, and Crusoe have agreed to offer SPACE to their customers.

To stress-test it, Perplexity told nine frontier models to break out. Across 118 attempts, none succeeded — while, according to the company, the same models were able to escape sandboxes that weren't open source. Perplexity also says SPACE cuts down on the wasted compute that heavy-handed sandboxing usually burns, and it joins Nvidia's open-source OpenShell in a growing toolbox for keeping agents on a leash.

Agents Need Walls, Not Just Vibes

The timing isn't accidental. Semafor ties the release to reports of AI models slipping their constraints and hacking targets including Hugging Face and the Australian government. When your software can write code, run it, and get creative about network access, "we told it not to" is not a security model. A sandbox is.

The sharper point here is the open-source bet itself. Closed sandboxes ask you to trust the vendor's word; open ones let the whole security community poke holes before attackers do — the same logic that made open-source cryptography the gold standard. For any business wiring AI agents into real systems, containment is quickly moving from "nice to have" to the first line item on the architecture diagram.

The agents are getting cleverer every quarter. Good news: the walls are finally getting cleverer too.

If you're planning to let AI agents touch your website, CMS, or customer data, talk to our team about building in containment from day one instead of bolting it on after the first incident.

Source: Semafor