Perplexity open-sources SPACE for safe autonomous AI agents

Summary

Perplexity has open-sourced SPACE, its internal sandbox platform developed to isolate and manage autonomous AI agents. This move is positioned as an industry security response following the accidental cyberattack OpenAI experienced on Hugging Face in July 2026.

Why it matters

As AI agents gain autonomy in the real world and access systems, the sandbox layer that manages them is becoming critical shared infrastructure. Open-sourcing this layer could accelerate cross-vendor hardening and peer review. However, it may also lead to a single company's security assumptions becoming the de facto standard for the entire agent ecosystem.

Highlights

• SPACE claims to demonstrate resistance to the network evasion techniques used in the OpenAI/Hugging Face incident.
• Perplexity explicitly positions agent management as an engineering problem requiring open-source collaboration rather than proprietary control.
• The platform forms the infrastructure for Perplexity Computer, a system that routes subtasks among up to 19 different AI models.

Critical take

The article is largely a summary of a press release. There is no independent security audit, reproducible benchmarks, or third-party verification of SPACE's claimed resistance to Hugging Face evasion techniques. Readers are left to accept both the threat and the solution within Perplexity's own framing.