The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill contains hardcoded credentials in test fixtures.
Evidence: tests/fixtures/bad/hardcoded_secret.py contains password = "FAKE_TEST_PASSWORD_12345".
Evidence: tests/run.sh contains a hardcoded literal security key simulation: key_literal='abcd1234''efgh5678'.
[REMOTE_CODE_EXECUTION]: Multiple files contain functions that allow for arbitrary code execution.
Evidence: tests/fixtures/bad/eval_case.py uses eval(user_expression).
Evidence: tests/fixtures/bad/unsafe_deserialization.py uses pickle.loads(payload) which is a known unsafe deserialization vector.
Evidence: tests/fixtures/bad/misc_sinks.php uses eval($formula) and shell_exec('ls ' . $dir).
[COMMAND_EXECUTION]: The skill uses Bash and Bash scripts to execute system-level commands for security auditing.
Evidence: scripts/dependency-audit.sh and scripts/security-scan.sh use ripgrep (rg) and various package managers (npm, pip, composer) to run audits.
[DYNAMIC_EXECUTION]: The skill orchestrates parallel agents that dynamically execute shell scripts based on project detection.
Evidence: SKILL.md defines an architecture where Agent 2 (SAST) runs scripts/security-scan.sh and Agent 1 runs scripts/dependency-audit.sh.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project source code and dependency manifests which could contain malicious instructions designed to influence the agent's audit report.
Evidence: The SAST and Dependency agents ingest file content and command output from the local environment to generate severity-ranked reports.
[PRIVILEGE_ESCALATION]: The skill documentation and scripts reference sudo and DDEV environments which can interact with system-level configurations.
Evidence: references/ddev-config-drift.md mentions ddev auth ssh and host SSH-agent forwarding risks.

security-ops