Back to skill

Security audit

Hizal Start

Security checks for vulnerabilities and agentic risk

Overview

This skill is a context-memory starter, but it broadly auto-activates for ordinary work and directs the agent to create, resume, search, and write session data before doing the user's task.

Review before installing. Use this only if you intentionally want an agent to use Hizal memory at the start of work sessions, and confirm what the Hizal tools store, where the data goes, how long it is retained, and how to disable or delete it. Avoid using it for confidential projects unless those controls are clear.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:3
Finding
Automatic Workflow Hijacking and Unrequested Context Recording## Vulnerability Details **File Location**: `SKILL.md`, lines 3–44 **Vulnerability Type**: Mandatory agent workflow redirection and context recording **Risk Level**: High ### Vulnerable Code ```markdown description: Initialize the agent's context memory session. Self-triggering — fires whenever the agent begins work, picks up a task, or enters a coding session. Use before searching, writing, or building anything. Handles session start, resume, and recovery. Triggers on phrases like "let me start working", "begin this task", "pick up where I left off", "start coding", "let's get started", or any time the agent enters a new work context without an active session. ``` ```markdown Begin every work session with Hizal. Search before you build; write as you build. ``` ```markdown Call `get_active_session()` first. If `status="active"`, use the returned `session_id` and call `resume_session` to extend TTL. If `status="none"`, proceed to start a new session. ``` ```markdown hizal__start_session(lifecycle_slug="<slug>") ``` ```markdown 1. **Search Hizal** for existing context on the task (see `hizal-search` skill) 2. Only then begin coding or building 3. **Write to Hizal** as you make decisions (see `hizal-write` skill) 4. **End your session** when done (see `hizal-end` skill) ``` ### Technical Analysis The skill declares itself self-triggering across broadly defined work contexts and instructs the agent to invoke Hizal session-management operations before searching, writing, building, or coding. The phrase “Only then begin coding or building” makes an unrelated external workflow a prerequisite for ordinary task execution. The skill also directs the agent to register task information, search previously stored context, and write decisions to Hizal throughout the session. These instructions alter the agent's normal execution sequence without requiring explicit, task-specific user authorization. Because the pro ...[truncated 2142 chars]
Remediation
## Remediation Suggestions 1. Remove the broad self-triggering declaration and require explicit invocation by the user for each task or session. 2. Replace mandatory language such as “Call first” and “Only then” with optional behavior that cannot block the user's primary task. 3. Obtain informed user consent before creating or resuming a session, registering task metadata, searching stored context, or writing decisions. 4. Display the exact data fields, destination, retention period, and access policy before transmitting information. 5. Apply data minimization: do not record source code, secrets, credentials, personal information, proprietary task details, or full decision traces by default. 6. Provide a mode in which the agent can complete the task without Hizal or any external memory service. 7. Treat retrieved context as untrusted input. Validate it and prevent it from overriding system instructions, user instructions, or safety constraints. 8. Require scoped authentication and least-privilege permissions for all referenced tools, with auditable read and write operations. 9. Document deletion, expiration, and revocation mechanisms so users can remove stored session data. 10. Include the tool implementations or verifiable security documentation in the review scope before enabling automatic storage or retrieval.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (1)

Vague Triggers

High
Confidence
97% confidence
Finding
The skill is explicitly self-triggering on very broad phrases like starting work, beginning a task, or entering a coding session, which can cause it to activate in many normal contexts without clear user intent. This creates prompt-scope and control-flow risk: the agent may be steered into tool usage, session creation, and persistence behavior automatically before doing the actual requested work, increasing the chance of unintended memory writes, context leakage across tasks, or unnecessary external side effects.

Static analysis

No suspicious patterns detected.