Comment by ollin

If the HuggingFace repo the agent broke into contains reference solution scripts for ExploitGym (i.e. for exploiting the vulnerabilities in the intended way), the agent can then run that reference code inside its original sandbox to retrieve the dynamically-generated flags.