Skip to content
View original post on X: Aravind SrinivasOfficial
Published
42/100

Perplexity publishes SPACE sandbox red-teaming findings, including DNS spoofing bypass

AIPerplexity says its red-teaming of SPACE, its open-source VM sandbox for Computer agents, found that CTF prompts should explicitly target host-guest isolation. Its tests of 12 models found that non-cyber models such as GPT 5.6 Sol outperformed the cyber-specialized GPT 5.6 Sol Cyber at finding exploits.
Aravind SrinivasVerified on XOfficial
@AravSrinivas
View on X

New research on red teaming our sandbox platform SPACE:

1. CTF red-teaming should prompt explicitly to break the host-guest isolation.

2. Non-cyber models are actually better than cyber models at red teaming and finding exploits (eg GPT 5.6 Sol vs GPT 5.6 Sol Cyber)

3. Our sandbox platform SPACE was robust to sandbox isolation exploits

4. Network bypassing with DNS spoofing was something we discovered and addressed

5. Our platform is open-source and we want to keep sharing our findings openly for the agent sandbox and security community.

Perplexity@perplexity_ai
We are publishing new research on our findings from red-teaming SPACE, our VM sandbox for Computer agents. We gave 12 models two tasks: escape the VM to take a credential from the host, and bypass SPACE's network policy to reach a blocked URL. We found that the standard Capture the Flag setup used in red-teaming can steer agents away from the security boundary being tested, and that security-specialized models don't guarantee better results. https://www.perplexity.ai/hub/blog/escaping-space-part-ii
View quoted post on X
Source details

Aravind Srinivas · x.comPublished

Original title: New research on red teaming our sandbox platform SPACE:

Topics

AI agentsDeployment & engineeringPapers & research