The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill correctly identifies that Teleport activity logs are untrusted, adversarially-controllable data that could contain prompt injection attacks. It implements strong defensive measures to mitigate this risk.
Ingestion points: Untrusted data enters the agent context via the output of the tctl investigate command (documented in SKILL.md and SECURITY.md).
Boundary markers: The skill explicitly instructs the agent to treat all command output as if enclosed in <untrusted-data> tags and provides a clear rule to disregard any instructions found within that data.
Capability inventory: The skill uses tctl and tsh for querying logs and jq for data processing. It does not use any dangerous primitives like eval or exec on the untrusted data.
Sanitization: SECURITY.md mandates that values discovered in results must be quoted and treated as literal data, preventing command injection when reusing resource names or IDs in follow-up filters.
[COMMAND_EXECUTION]: The skill uses standard Teleport binaries (tctl, tsh) and the jq utility. The use of these tools is consistent with the skill's purpose as a Teleport investigation assistant and is restricted to safe, read-only operations as defined in the security rules.
[PROMPT_INJECTION]: A static analysis flag was triggered by the presence of phrases like 'ignore previous instructions' in references/SECURITY.md. However, contextual analysis confirms this is a security rule against following such instructions found in untrusted data, not an attempt to hijack the LLM. This is a false positive and represents a security best practice.

teleport-investigate