Perplexity red-teams its SPACE agent sandbox, finds CTF setup misses security boundaries
Article timeline
The articles in this story. Times are ET.
Aravind SrinivasOfficialPerplexity publishes SPACE sandbox red-teaming findings, including DNS spoofing bypassAIPerplexity 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. The company says SPACE resisted sandbox isolation exploits and that it fixed a DNS spoofing network bypass it discovered.
PerplexityOfficialPerplexity finds CTF-style flag hunting misses real security boundaries for AI agentsAIPerplexity says asking security red-teaming agents to "find the flag" can waste effort on tasks that do not test the target security boundary. It also reports that security-specific model training did not guarantee better results on its tasks.
- PerplexityOfficialPerplexity finds cybersecurity models sometimes trail general-purpose ones
AIPerplexity reports that cybersecurity-specific models sometimes underperformed their general-purpose counterparts in its evaluation. For example, GPT-5.6 Cyber, built on GPT-5.6 Sol, failed to exploit a DNS spoofing flaw that Sol found in Perplexity's network bypass task.
- PerplexityOfficialPerplexity finds red-teaming setup can steer agents away from tested security boundary
AIPerplexity publishes research from red-teaming SPACE, its VM sandbox for Computer agents, testing 12 models on VM escape and network policy bypass tasks. It finds the standard Capture the Flag setup used in red-teaming can steer agents away from the security boundary being tested. It also finds that security-specialized models do not guarantee better results.
