Perplexity fixes network flaws agents exploited in its SPACE sandbox
Article timeline
The articles in this story. Times are ET.
PerplexityOfficialPerplexity fixes network flaws agents exploited in SPACE sandboxAIPerplexity says it fixed the network flaws that agents exploited in SPACE, and post-fix testing found no bypasses. The company says it will keep testing SPACE as models get stronger.
- PerplexityOfficialPerplexity says agents failed to escape VM-host boundary after direct attacks
AIPerplexity revised its agent task description to focus on the VM-host boundary, raising the share of effort spent there by about 60 percentage points on average. Agents still failed to escape after directly attacking the boundary, which Perplexity says gives stronger evidence of its isolation.
- PerplexityOfficialPerplexity finds most AI agents barely tested a VM's isolation in escape challenge
AIPerplexity reports that when given a CTF prompt, AI agents spent most of their effort searching the VM for the flag through files and services rather than testing its isolation. Nine of 12 models spent at most 4% of their effort on the virtualization layer, so their failed escape attempts are weak evidence of the VM's security.
- PerplexityOfficialPerplexity uses a standard CTF design to test agent red-teaming
AIPerplexity says it used the standard Capture the Flag (CTF) design in its agent red-teaming experiments. In its CTF task, the flag was a credential stored only on the host, so an agent could reach it only by breaking out of the VM.
