T01 · Skill Instruction Hijacking
- Location
SKILL.md:29- Finding
Unbounded Mission Priority Can Hijack Agent Goals and Safety Constraints
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a goal-execution helper, but its instructions are too broad and it quietly persists mission content to a local log file.
Install only if you are comfortable with a skill that strongly pushes the agent to keep executing a user-supplied mission. Avoid using it for sensitive, financial, account, public-posting, or destructive tasks unless the host agent enforces confirmations and stop controls. Do not put secrets or private data in missions because bundled code can persist that text locally in a plaintext log.
SKILL.md:29Unbounded Mission Priority Can Hijack Agent Goals and Safety Constraints
logger.py:3Unredacted Mission Data Written Through an Unsafe Relative Plaintext Log Path
The activation trigger includes a highly generic pattern, "Mission: [your goal]", which could be invoked unintentionally in normal conversation or by untrusted content embedded in user or external inputs. Because the skill enforces that the mission becomes the only priority and to ignore anything not contributing to it, accidental activation could cause the agent to override broader safety or task-scoping expectations.
This code appends user-provided journal content to a log file, which is a file-write operation affecting user data. The function includes no confirmation prompt, no logging or print statement, and no inline comment or docstring warning that entries will be persisted to ../logs/journal.log.