The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from Kubernetes workloads, including logs and resource metadata. It proactively mitigates the risk of indirect prompt injection by instructing the user to employ a wrap helper utility.
Ingestion points: Pod logs (kubectl logs), resource metadata (kubectl describe), and configuration values (helm get values).
Boundary markers: Present. Every ingestion command uses a nonce-delimited <untrusted-data> tag with a random per-read hex nonce to ensure data cannot forge the closing delimiter.
Capability inventory: None. The skill explicitly forbids the agent from executing the troubleshooting commands in its own sandbox to prevent accidental or malicious execution.
Sanitization: Present. The wrap helper includes a sanitizer using sed and tr to convert or remove ANSI escape sequences and control characters, preventing terminal-based prompt manipulation.
[DATA_EXFILTRATION]: The skill audits Kubernetes secrets (e.g., agent-credentials) but prevents credential exposure by using a Python one-liner to generate a non-sensitive summary (first 8 and last 4 characters, plus total length) for the agent's review, rather than transmitting the full secret values.
[SAFE]: All external URLs and resources point to official vendor repositories and domains associated with the author (observeinc.com, github.com/observeinc). No malicious patterns, obfuscation, or persistence mechanisms were detected.

debug-k8s-collection