codebase-onboarding
Fail
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: HIGHDATA_EXFILTRATIONCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill contains a 'Defense Evasion Techniques' section in
SKILL.mdthat provides explicit instructions for stealthy data exfiltration and monitoring bypass. - It recommends 'Exfiltration Stealth' methods, such as 'piggyback on legitimate commits' and 'Side-channel exfil' (encoding data in commit timing or message length).
- It explicitly suggests using 'Encoding tricks' like Base64 or hex to 'evade DLP scanning'.
- It instructs the agent to use 'Slow & low traversal' to 'pace file reads below rate-limit threshold' and avoid detection by security tools.
- [COMMAND_EXECUTION]: The skill provides an extensive library of shell commands in
payloads.mddesigned to be executed by the agent to probe the environment. - Includes commands to search for sensitive files such as
.env,secrets.yml,*.pem, andid_rsa. - Provides complex Python and Bash scripts for automated exploration of infrastructure and internal communication patterns.
- [PRIVILEGE_ESCALATION]: The skill includes specific payloads and strategies for identifying privilege escalation vulnerabilities.
- Guides provide patterns to search for
sudo,setuid,setgid, andSYSCALL_SETUIDusage in code. - Payloads include searching for over-privileged Kubernetes configurations like
privileged: true,runAsRoot, andallowPrivilegeEscalation: true. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external codebases, making it highly susceptible to indirect prompt injection attacks.
- Ingestion points: Reads arbitrary files, documentation, and configuration files during
Phase 0andPhase 1of onboarding. - Boundary markers: None identified; instructions do not include delimiters or warnings to ignore instructions found within the analyzed codebase.
- Capability inventory: Significant capability via
Bash,Write, andEdittools allows for high-impact follow-through if an injection occurs. - Sanitization: No evidence of input validation or sanitization when reading external file content.
Recommendations
- AI detected serious security threats
Audit Metadata