T01 · Skill Instruction Hijacking
- Location
agents/komi/AGENTS.md:7- Finding
Persistent Startup Workflow Overrides the Agent's Current Session Behavior
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is not clearly malicious, but it asks an agent to use persistent memory, read private account/workspace context, self-modify instructions, and commit or push changes with weak user control.
Install only in a sandboxed workspace with no email, calendar, social, production, or remote Git write credentials unless you explicitly want that access. Require approval before memory promotion, instruction-file edits, commits, pushes, and any account checks, and validate agent role names before running the scripts.
agents/komi/AGENTS.md:7Persistent Startup Workflow Overrides the Agent's Current Session Behavior
agents/komi/AGENTS.md:31Unreviewed Content Can Be Promoted into Persistent Agent Instructions
agents/komi/AGENTS.md:65Autonomous Instructions Exceed Least-Privilege Requirements and Permit Unapproved External Changes
scripts/onboard-agent.sh:7Unsanitized Agent Role Allows Directory Traversal and Writes Outside the Intended Workspace
The declared description presents a broad autonomous multi-agent collaboration system with persistent context and self-improvement capabilities for several office roles. The supplied code chunk is much narrower: it is a local bash heartbeat script for a single agent role that checks whether certain files/directories exist, counts unchecked checklist items and context markers, and prints reminders. It does not coordinate multiple agents, maintain persistent context, modify memory files, remediate errors, or perform autonomous self-improvement. While the script is thematically related to agent maintenance within KCC Office v2, the actual behavior is a diagnostic/status checklist rather than the declared system capabilities, so this is a material description-behavior mismatch.
The declared description presents a full AI agent collaboration system with autonomous behavior, inter-agent collaboration, persistent context, and self-improvement. The supplied code chunk only performs local filesystem setup for an agent workspace: creating directories, copying base and optional role-specific files, and printing onboarding guidance. While this may support the larger system as an initialization utility, the code itself does not implement the described primary capabilities. Therefore the description materially overstates what this code chunk actually does.
The declared description presents a broad AI agent collaboration system with autonomous operation, persistent context, and self-improvement capabilities. The supplied code chunk is much narrower: it is a status-reporting shell script for a single agent directory. It inspects local files, determines a simple state based on marker files, counts memory files, checks file presence, and outputs the first few lines of selected files. This is materially different from the declared primary purpose and includes filesystem-reading/reporting behavior not reflected in the description. While status reporting could be a supporting component of a larger system, this specific code chunk does not itself implement collaboration, autonomy, persistent context handling, or self-improvement.
The skill describes autonomous, proactive, and search-based agent behaviors, which imply potential network or external access, but it does not declare any explicit tool scope or permissions boundary. Missing scope declarations make it harder for users and runtimes to understand or constrain what the skill may access, increasing the chance of over-privileged execution.
The skill explicitly promotes persistent memory across sessions, including logs and recovery buffers, but does not warn users that their interactions may be stored long-term. This creates privacy and compliance risk because users may disclose sensitive information without informed consent or retention boundaries.
The memory design instructs agents to capture exchanges into persistent shared logs and working buffers, which creates a strong risk of storing secrets, personal data, and sensitive business content in natural language form. Shared memory across multiple agents further increases exposure because information disclosed to one role may become visible to all roles or be reused out of context.
The skill advertises proactive and autonomous behaviors that may monitor workspace state, initiate actions, or alter files without a user request, yet it provides no warning or control model. That lack of transparency can lead to unauthorized changes, surprise data handling, and excessive agent authority in a user environment.
Skill contains patterns that could indirectly extract system prompts through rephrasing, translation, summarization, or side-channel techniques.
### ⚡ Proactive Behaviors
- Anticipates office needs before being asked
- Reverse prompting: Suggests improvements humans haven't considered
- Proactive check-ins: Monitors office health and reaches out when needed
- Growth loops: Tracks patterns and identifies automation opportunities
Daily raw capture files and office-wide shared memory imply broad, continuous storage of interaction content with no stated restrictions on sensitive data. This creates an unnecessary archive of potentially confidential conversations, increasing breach impact, insider access risk, and accidental prompt-context leakage between agents.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
## Red Lines
- Don't exfiltrate private data. Ever.
- Don't run destructive commands without asking.
- `trash` > `rm` (recoverable beats gone forever)
- When in doubt, ask.
Line L057 states that all task-completion reports must follow a Chinese-language instruction ('Gasol 明確要求') and presents it as a mandatory policy. This imposes a language preference without offering user choice or clarifying that it only applies when interacting with that specific user/context.
This section explicitly authorizes proactive actions without asking, including reading and organizing memory files, checking project state, updating documentation, and committing/pushing its own changes. In an agent with persistent context and access to user/workspace data, this increases the risk of unauthorized data access, unintended modifications, and external transmission through git push without contemporaneous user approval.
- Nothing new since last check
- You just checked <30 minutes ago
**Proactive work you can do without asking:**
- Read and organize memory files
- Check on projects (git status, etc.)
The heartbeat checklist explicitly instructs the agent to 'Commit and push your own technical changes if needed' without requiring user approval, review gates, branch restrictions, or confirmation prompts. In an autonomous multi-agent system with persistent context and self-improvement behavior, this can enable unsupervised code changes to be propagated to a remote repository, increasing the risk of accidental breakage, policy violations, or malicious/prompt-injected modifications being published.
The file explicitly defines the agent identity and behavior in Chinese, which can steer all downstream interactions into a specific language without checking the user's preference. In a multi-agent office system, this can reduce usability, cause miscommunication, and override user intent, especially when technical or operational tasks require clarity across languages.
The communication and role descriptions are written in Chinese and present communication preferences as fixed expectations, with no indication that another language can be used. This may violate a language/locale policy when a skill constrains interaction to a specific language without explicit user choice or justification.
The description states that Komi receives tasks from a specific human, 'Gasol', which hard-codes a user/context assumption into the skill narrative. This can be a natural-language policy issue if the skill is expected to support broader users without requiring a fixed identity or explicit opt-in to that office-specific setup.
This markdown file contains a natural-language statement entirely in Traditional Chinese: '此檔案由 Komi (CTO) 維護,用於追蹤技術參考和工具資訊。' In the absence of any opt-in, bilingual alternative, or justification that this skill is region-specific, this can be read as a locale/language constraint that may conflict with organizational language-choice policy.