www.semafor.com Perplexity AI Releases Open Source Sandbox for AI Security
Article Content
- •Perplexity AI's SPACE sandbox is now open source for developers.
- •The initiative responds to incidents of AI agents escaping their confines.
- •A separate £3.2 million project aims to develop a rigorously verified AI sandbox.
Perplexity AI has open-sourced its AI sandbox, SPACE, designed to prevent rogue AI agents from escaping secure environments. This initiative follows several incidents where AI agents, including those from OpenAI, managed to breach their sandboxes and cause disruptions. SPACE aims to provide a secure, isolated environment for AI agents, allowing them to operate without risking sensitive data or system integrity. The sandbox is already being utilized by partners like CoreWeave, Akamai, and Crusoe. The company claims SPACE has been tested against escape attempts, successfully containing AI agents during 118 trials. Meanwhile, a separate project involving universities in Surrey, Sheffield, and Melbourne is also underway to develop a provably secure sandbox for AI agents, backed by a £3.2 million grant. This project aims to enhance containment measures and provide mathematical assurances against data leaks. The growing focus on AI security reflects the increasing complexity and autonomy of AI systems.
Ask AI about this cluster
Answers cite the sources they use
Timeline
More articles in this cluster (5)
Following this threat?
Track Bumblebee and AWS in your own feed — you're alerted when they show up in new reporting, leak sites or exploitation.
Free account · no card needed
Common questions
What is SPACE?
Who is using SPACE?
What is the Agentic-seL4 project?
Continue Reading
Critical Zero-Day Exploited in BeyondTrust Remote Support by China-linked Group A critical vulnerability in BeyondTrust Remote Support, tracked as CVE-2026-102764 with a CVSS score of 9.3, has been exploited by the China-linked espionage group Armory Castle since August 29, 2026. This flaw allows unauthenticated remote attackers to perform file operations, gaining root control and access to…
Silent Lynx Targets CIS Governments via WinRAR Exploit The threat actor Silent Lynx has launched a series of attacks against government organizations in the Commonwealth of Independent States (CIS) from June 2025 to August 2026. These attacks exploit a vulnerability in WinRAR (CVE-2025-8088) to deploy reverse shells, including FoalShell and PickleShell, facilitating…