SPACE Sandbox Red-Teaming Finds Non-Cyber Models Better at Exploits
New red-teaming research on the SPACE sandbox platform found that non-cyber models outperform cyber-specialized models at finding exploits, and that CTF exercises should explicitly prompt for host-guest isolation breaks. SPACE proved robust to sandbox isolation exploits, while DNS spoofing network bypassing was discovered and addressed. The open-source platform's findings are being shared with the agent sandbox and security community.
Event version 1.
Key insights
Non-cyber models outperform cyber models at red teaming
Non-cyber models are better than cyber models at red teaming and finding exploits. GPT 5.6 Sol outperformed GPT 5.6 Sol Cyber in this role.
CTF red-teaming should target host-guest isolation
CTF red-teaming exercises should explicitly prompt participants to break the host-guest isolation.
SPACE resisted sandbox isolation exploits
The SPACE sandbox platform was robust to sandbox isolation exploits during red-teaming.
DNS spoofing bypass discovered and fixed
Network bypassing via DNS spoofing was discovered during testing and has been addressed.
Open-source findings shared with security community
SPACE is open-source, and its findings are being shared openly for the agent sandbox and security community.