The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill instructs agents to query application logs, which are an ingestion point for untrusted data. Malicious users can inject instructions into logs (e.g., via User-Agent strings or request parameters) that may influence an agent when it parses them during debugging.
Ingestion points: Agents read logs/app.json through shell tools or a development REST endpoint.
Boundary markers: Absent; suggested documentation does not provide delimiters to isolate log content from system instructions.
Capability inventory: Agents are permitted to execute shell commands (grep, tail, jq) and perform network queries.
Sanitization: While the skill emphasizes redacting PII for privacy, it does not provide mechanisms to sanitize log entries for prompt injection patterns.
[Data Exposure & Exfiltration] (SAFE): The skill recommends an unauthenticated /__dev/logs endpoint for development convenience. This is mitigated by explicit requirements to disable the endpoint in production and instructions to redact sensitive fields like tokens and passwords.
[Unverifiable Dependencies & Remote Code Execution] (LOW): The skill suggests installing reputable and standard libraries such as pino, structlog, and lograge. These are considered low risk under the trusted source guidelines.

agent-telemetry