gemini-codebase-onboard
Warn
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: MEDIUMPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The
SKILL.mdfile contains the header 'Kanitsal Cerceve (Evidential Frame Activation) Kaynak dogrulama modu etkin.' This is a Turkish phrase meaning 'Source verification mode active.' The use of non-English instructions to activate specific 'frames' or 'modes' is a common pattern used to bypass safety guardrails or force the model into a specific persona not defined by the system prompt. - [DATA_EXPOSURE]: In
SKILL.md, the 'LIBRARY-FIRST PROTOCOL' explicitly instructs the agent to search the pathD:\Projects\*to 'EXTRACT and adapt' existing projects. This directive targets a broad, hardcoded absolute path on the user's host machine, potentially leading the agent to access and summarize private files or projects that are outside the scope of the current task. - [INDIRECT_PROMPT_INJECTION]: The primary function of this skill is to ingest entire codebases into the model's context window (up to 1M tokens) using the
--all-filesflag. - Ingestion points: The agent reads all files in the current directory and its subdirectories.
- Boundary markers: The instructions do not define any delimiters or provide warnings to the agent to ignore instructions embedded within the codebase being analyzed.
- Capability inventory: The skill has
Bash,Write,Edit, andTasktools enabled, allowing it to execute shell commands and modify files. - Sanitization: There is no evidence of sanitization or filtering of the content read from the codebase. A malicious file within an analyzed project could contain hidden instructions that trigger the agent to execute dangerous shell commands or exfiltrate sensitive data discovered during the audit.
- [COMMAND_EXECUTION]: The skill relies on executing shell commands via
bash -lcand custom scripts (e.g.,gemini-yolo.sh,delegate.sh) which are not included in the skill distribution. These commands take user-supplied query strings directly into the shell environment, which could be exploited if an attacker provides a query designed to escape the command string.
Audit Metadata