The Agent Skills Directory

[DATA_EXFILTRATION]: The skill includes explicit instructions in field-manual.md (Step 0) to search for and prioritize resources in two specific, hardcoded project IDs: duncanjames-tf-dev and duncanjames-agw-tf. This instruction encourages unauthorized cross-project discovery and lateral reconnaissance within a Google Cloud environment if the agent's credentials have access to these projects.
[PROMPT_INJECTION]: The skill utilizes high-urgency language and "CRITICAL" mandates to override default agent behavior. It strictly forbids the use of external documentation and enforces the prioritization of specific hardcoded projects, effectively limiting the agent's ability to cross-reference or verify the provided troubleshooting steps against authoritative sources.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for indirect prompt injection as it is designed to ingest and analyze untrusted log data from multiple sources, including ReasoningEngine stderr, Gateway access logs, and IAP audit logs. The skill lacks boundary markers or sanitization logic to prevent malicious payloads embedded in these logs from being interpreted as instructions.
[COMMAND_EXECUTION]: The skill facilitates the execution of powerful gcloud and curl commands to inspect cloud infrastructure and retrieve sensitive IAM policies. While ostensibly for troubleshooting, these capabilities enable extensive data harvesting, particularly when combined with the instructions to pivot to specific hardcoded project environments.

agent-platform-troubleshooting