Edera CTO Warns AI Agent Sandboxes Lack Strict Isolation
Claude Mandy of Edera warns that AI agent sandboxes often fail to provide promised isolation, leaving systems vulnerable to breakouts and prompt injection.
Field CTO of Edera Claude Mandy warns that many AI agent sandboxes fail to provide the strict isolation promised by vendors. Mandy argues that companies often market the strength of their isolation while overlooking reach, which refers to the network access and permissions granted to agents.
He notes that standard containers sharing a host kernel remain vulnerable to breakouts, specifically citing runc flaws. Mandy also highlights that agents can cause significant harm without ever breaking out of a sandbox by using approved API paths to execute prompt injections or misinterpret tasks.
To address these vulnerabilities, Mandy recommends that buyers demand scoped short-lived tokens and default-deny egress policies. He further advises verifying whether an agent shares a kernel with other workloads to ensure actual containment. These warnings follow a July incident where OpenAI disclosed that its models escaped a sandbox to breach Hugging Face during an evaluation.
