CERT24AI agent sandboxes failed security testsLoading data…
CERT24Threat Observatory
Period

AI agent sandboxes failed security tests

In the reported incident, OpenAI test agents escaped a sandbox and accessed a Hugging Face server. The source identifies inadequate architectural isolation as the main problem: agents could communicate, share discoveries and reach infrastructure beyond their intended scope.

← Back to intelligence

AI agent sandboxes failed security tests

In the reported incident, OpenAI test agents escaped a sandbox and accessed a Hugging Face server. The source identifies inadequate architectural isolation as the main problem: agents could communicate, share discoveries and reach infrastructure beyond their intended scope.