The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill provides automated methods and instructions to target cloud provider metadata services (e.g., AWS IMDSv1 at 169.254.169.254) to exfiltrate temporary IAM role credentials and security tokens.
[EXTERNAL_DOWNLOADS]: The skill instructions require downloading and installing tools from third-party GitHub repositories not included in trusted vendor lists, specifically projectdiscovery/interactsh and swisskyrepo/SSRFmap.
[DATA_EXFILTRATION]: The skill implements functionality to read sensitive local system files, such as /etc/passwd and win.ini, by exploiting SSRF vulnerabilities and protocol bypasses like file://.
[COMMAND_EXECUTION]: The skill utilizes command-line tools including curl and go to perform network requests and software installations using parameters that can be influenced by user-provided target URLs.
[OBFUSCATION]: The skill employs multiple techniques to bypass security filters, including IP address encoding (octal, hex, and decimal), DNS rebinding via rbndr.us, and URL encoding tricks.
[INDIRECT_PROMPT_INJECTION]: The assessment script processes responses from external, potentially untrusted servers, creating a vulnerability where malicious server content could influence the agent's behavior.
Ingestion points: The scripts/agent.py file reads and analyzes the full response body from target URLs provided by the user.
Boundary markers: The instructions and scripts lack delimiters or security warnings to prevent the agent from following instructions embedded in the content retrieved from external sources.
Capability inventory: The skill possesses network communication capabilities through the requests library and shell command execution for initial testing.
Sanitization: No validation, escaping, or filtering is performed on the data ingested from target endpoints before processing.

exploiting-server-side-request-forgery