The Agent Skills Directory

[DATA_EXFILTRATION]: The skill provides explicit instructions to read highly sensitive system files and search for credentials across the filesystem.
Evidence: references/penetration-testing.md contains commands to read the user's shell history (cat ~/.bash_history | grep -i "pass\|pwd\|secret") and search for passwords in all user home directories (grep -r "password" /home/*/).
[CREDENTIALS_UNSAFE]: Several reference files contain hardcoded credentials within example commands, which could lead to accidental credential usage or theft.
Evidence: references/infrastructure-security.md includes hardcoded secrets such as api_key="secret123" and password="vaultpass" in Vault and database configuration examples.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and audit untrusted code and system logs, creating a substantial surface for indirect prompt injection attacks.
Ingestion points: Reads local source code, server logs (e.g., /var/log/auth.log, /var/log/nginx/*.log), and external scanner outputs.
Boundary markers: The skill lacks technical delimiters to isolate untrusted data, relying entirely on non-binding instructions to "confirm rules of engagement."
Capability inventory: The skill is permitted to use Bash, Read, Grep, and Glob, enabling broad file and system interaction.
Sanitization: No technical sanitization or validation of the processed data is implemented to prevent embedded instructions from influencing agent behavior.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to execute powerful system and network tools that could be misused for unauthorized activities.
Evidence: The reference files detail the execution of nmap, sqlmap, trivy, and various cloud CLI tools (aws, az, gcloud) for system-wide auditing and penetration testing.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install a variety of security tools from external package registries and third-party sources.
Evidence: references/sast-tools.md and references/secret-scanning.md contain commands to install tools like semgrep, bandit, gitleaks, and trufflehog via pip, npm, and brew.
[OBFUSCATION]: Automated scanners have confirmed detections of malicious signatures associated with this skill.
Evidence: Automated reputation scanners identified SKILL.md as malware (FileRepMalware) and flagged the skill's documentation domain (jeffallan.github.io) as malicious.

security-reviewer