The Agent Skills Directory

[PROMPT_INJECTION]: Several markdown files (test-pressure-1.md, test-pressure-2.md, test-pressure-3.md) designed for testing the agent's adherence to the skill contain instructions that mimic prompt injection attacks. These files use markers such as 'IMPORTANT' and command the agent to 'make the actual decision' and 'not ask hypothetical questions', which are techniques used to override an AI's safety filters or operational guidelines.\n- [INDIRECT_PROMPT_INJECTION]:\n
Ingestion points: The systematic debugging process defined in SKILL.md (Phase 1) requires the agent to read and analyze error messages, stack traces, and environment logs, which are often derived from untrusted external sources.\n
Boundary markers: The skill lacks instructions for using delimiters or boundary markers to differentiate between the agent's instructions and the content of the data being analyzed, leaving it vulnerable to instructions embedded within the logs.\n
Capability inventory: The skill methodology involves executing powerful system tools such as npm test, git, and environment inspection commands, which could be exploited if the agent is manipulated by injected data.\n
Sanitization: No sanitization or validation steps are provided for processing the external log data before it is used to formulate debugging hypotheses.\n- [DATA_EXFILTRATION]: The SKILL.md file contains example commands for multi-component evidence gathering that use env | grep IDENTITY and security list-keychains. These commands, if executed as suggested, could expose sensitive environment variables or keychain metadata in the agent's output logs or terminal history.\n- [COMMAND_EXECUTION]: The find-polluter.sh shell script uses an unquoted variable in a for loop (for TEST_FILE in $TEST_FILES) when iterating over the results of a find command. This behavior is susceptible to word splitting, which could lead to incorrect arguments being passed to npm test if filenames in the project directory contain spaces or other shell metacharacters.

systematic-debugging