The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted Kubernetes pod traffic (L7 protocols like HTTP, gRPC, and DNS) which serves as a significant ingestion point for potentially malicious instructions embedded in network payloads.
Ingestion points: Network traffic captured via capture_traffic, list_l4_flows, and get_l4_flow_summary.
Boundary markers: None identified in the prompt templates to distinguish between captured traffic data and agent instructions.
Capability inventory: The skill can perform cluster-wide traffic captures and export data via export_pcap.
Sanitization: The skill includes a warning that captured traffic may contain PII or secrets, but no automated sanitization is implemented for the data presented to the agent.
[DATA_EXFILTRATION]: The export_pcap tool allows for the extraction of captured network traffic from the Kubernetes cluster. While this is the intended functionality of the tool, it creates a vector for moving potentially sensitive data (including decrypted TLS traffic and application-layer secrets) to external environments.
[COMMAND_EXECUTION]: The documentation references the use of kubectl and helm for cluster management and port-forwarding. While these are presented as setup instructions for the user, they involve high-privilege interaction with the Kubernetes API.

kubeshark-traffic