The Agent Skills Directory

[COMMAND_EXECUTION]: The instructions in references/replication.md include the command rm -rf /var/lib/postgresql/14/main/*. While intended for standby server setup, this is a highly destructive operation that will permanently delete all database files in the target directory if executed without extreme caution or in the wrong context.
[METADATA_POISONING]: Automated security scans have identified the documentation URL https://jeffallan.github.io/claude-skills/skills/infrastructure/postgres-pro/ in SKILL.md as malicious/blacklisted. Additionally, the file SKILL.md itself has been flagged for containing malicious patterns by reputation scanners.
[REMOTE_CODE_EXECUTION]: The skill recommends installing and running external utilities and extensions (e.g., apt-get install pg-auto-failover, pg_repack, pg_basebackup) from sources outside the immediate skill package, which involves executing code with system-level privileges.
[INDIRECT_PROMPT_INJECTION]: The skill frequently interacts with pg_stat_statements and other monitoring views that store raw query text from the database. This creates a surface for indirect prompt injection where malicious instructions embedded in database queries could influence the agent's behavior during performance analysis tasks.
Ingestion points: Reading from pg_stat_statements and pg_stat_activity (detected in SKILL.md, references/maintenance.md, references/performance.md).
Boundary markers: None identified; queries are processed as raw text for analysis.
Capability inventory: Subprocess execution via shell commands (rm, systemctl, pg_ctl), database write operations, and extension management.
Sanitization: No evidence of input validation or instruction filtering for retrieved database metadata.

postgres-pro