The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The script scripts/gremlin-setup.sh downloads a large ZIP archive (approximately 379MB) from a personal GitHub repository (https://github.com/nikhilbonte/gremlin-100/archive/refs/tags/gremlin-1.zip). This source is not associated with any verified or well-known organization and poses a significant supply chain risk.
[REMOTE_CODE_EXECUTION]: Software downloaded from an unverified third-party source is installed and executed on the atlas-0 pod within the Kubernetes environment. Furthermore, the gremlin-run.sh script explicitly allows the agent to copy and execute arbitrary local Groovy scripts to the pod, which can bypass the stated "read-only" purpose and lead to data modification or system compromise.
[COMMAND_EXECUTION]: The skill uses kubectl exec and kubectl cp to execute code and transfer files to a production or staging container. This pattern provides the AI agent with a pathway to execute arbitrary logic within the cluster's internal network.
[PROMPT_INJECTION]: The SKILL.md file contains a shell command injection vulnerability. It interpolates the $ARGUMENTS variable directly into a Bash execution string: tools/gremlin-atlas/scripts/gremlin-run.sh $ARGUMENTS. A malicious user could provide input containing shell metacharacters (e.g., ;, &, |) to execute unauthorized commands on the local system running the agent.

gremlin-atlas