Back to skill

Security audit

Secretary Skill

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed local goal-planning skill with storage and reminder features, but no evidence of deception, exfiltration, or destructive behavior.

Install only if you are comfortable having goals, progress, and review notes saved locally and referenced across sessions. Avoid storing secrets or highly sensitive health, financial, or workplace details unless you understand where the files and memory pointers are kept, and review any recurring reminders before enabling them.

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
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Rogue AgentSelf-Modification, Session Persistence
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (22)

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description promises a substantial multi-stage goal-management system, but the supplied code chunk contains only a placeholder comment and no functional behavior. Because the actual code does not demonstrate any of the declared capabilities, the description does not accurately represent what this code chunk actually does.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The code clearly supports de-ambiguation, vertical splitting, and cron automation, matching part of the description. However, the declared purpose emphasizes a full lifecycle system with five stages including persistent storage and ongoing review. In the supplied code, all state is in memory only, there is no database/file/API persistence, no retrieval or tracking subsystem, and no review workflow beyond generating reminder payload text. This is a material description-to-behavior mismatch because major declared stages and lifecycle capabilities are not present in the implementation.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The declared description promises a 5-stage end-to-end goal management pipeline, including de-ambiguation, splitting, persistent storage, automation, and review. The supplied code only covers the persistent storage/review-support portion: rendering plans to markdown, parsing them back, updating task checkboxes, and adding alignment log notes. It does not implement goal clarification, milestone/task decomposition, or any cron/automation behavior. This is a material scope mismatch because the actual code is only one subsystem of the declared full-lifecycle capability.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The declared description promises an end-to-end goal management pipeline with five stages, including clarification, decomposition, persistence, automation, and review. The actual code chunk only contains unit tests for a Markdown persistence layer. It demonstrates file creation, serialization/deserialization of goals/milestones/tasks, task status updates, alignment note insertion, and slugification. Persistent storage is represented, but the broader lifecycle features—especially de-ambiguation, splitting, and cron automation—are not present in this code. Because the supplied chunk’s actual behavior is materially narrower than the declared purpose, this is a mismatch.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
84% confidence
Finding

The skill explicitly instructs the agent to read and write files under ~/.hermes/goals, but it does not declare any scoped permissions or allowed-tools boundary. That creates an authorization ambiguity where storage-capable behavior is described without guardrails, making accidental overreach or unsafe tool exposure more likely.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill is designed to persist user goals, progress, metrics, and review notes to disk and memory, but it gives no warning that this information may include sensitive personal, financial, health, or career details. In a planning skill, such content is often highly personal, so silent persistence materially increases privacy exposure and surprise retention risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The de-ambiguation protocol mandates asking the user a specific Chinese phrase, which imposes a language choice without user opt-in. This is a natural-language policy issue because the file does not say the skill is Chinese-only or provide an alternative based on user preference.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
89% confidence
Finding

The skill intentionally creates session persistence by writing goal data to disk and saving a memory pointer for later reuse, but it does so without clear consent, retention limits, or deletion guidance. Persistent state is especially sensitive in a planning assistant because goals and progress logs can reveal durable personal patterns and priorities over time.

Content

Scanner excerpt · SKILL.md (reported line 123)May include surrounding context.

md
When the user says "存下这个目标" or the goal is locked and split:
1. Generate a unique slug from the title (lowercase, hyphens)
2. Build the full markdown file with frontmatter + body
3. Write to `~/.hermes/goals/<slug>.md`
4. Save to **memory** a brief pointer: "Goal '<title>' (slug) — ~/.hermes/goals/<slug>.md — status: active — target: <target> — progress: <current>"
5. Confirm to the user the goal has been filed

Ssd 3

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The cross-session memory pointer stores plain-language summaries of goals, targets, statuses, and progress for later recall and exposure. Because this skill handles long-term personal planning, those summaries can reveal sensitive life details across sessions without minimization or explicit consent.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill instructs the agent to read all stored goal files and display full contents on request, which can expose previously stored sensitive information far beyond what the user may currently expect. In this context, goal files may contain financial, health, career, timeline, and reflection data, so broad retrieval and display increases disclosure risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This code presents key prompts and interaction text in Chinese, including readiness checks that expect Chinese responses, while not offering any language selection or opt-in. That creates a natural-language policy concern because the skill imposes a locale/language on users rather than allowing choice.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The docstring says 'Show progress toward the final goal,' which implies a read/report operation. In practice, when a plan file is loaded, the command prompts for a task id and writes completion state back via update_task_status, making it a state-changing command as well as a status view.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The manifest says the skill provides a full lifecycle pipeline with de-ambiguation, vertical splitting, persistent storage, cron automation, and ongoing review. This file's top-level documentation and code only implement de-ambiguation, task splitting, and cron payload generation; there is no storage or review functionality in the actual code shown.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The question bank hard-codes all user-facing prompts in Chinese, which imposes a specific language on users with no visible choice or opt-in. This matches the locale/language policy violation category because the file provides no alternative language path or justification for a Chinese-only experience.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

save_plan accepts an arbitrary filepath and writes content there without constraining the destination to the intended plans directory. If an attacker can influence filepath, they could overwrite arbitrary user-accessible files, which is more concerning in an automation-oriented skill that persists and refreshes plans over time.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
82% confidence
Finding

The file hard-codes section names such as ## 校准记录 / Alignment Log and other bilingual template labels, implying a fixed locale/output format rather than user-selected language preferences. Under the stated policy, forcing a specific language or locale without opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Tainted flow: 'text' from pathlib.Path.read_text (line 104, file read) → pathlib.Path.write_text (file write)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · src/markdown_store.py (reported line 107)May include surrounding context.

python
text = text.replace(
                "\n---\n", f"\n{marker}\n\n{entry}\n\n---\n", 1
            )
            path.write_text(text, encoding="utf-8")

    # ── path resolution ────────────────────────────────────────────────

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill describes recurring cron-based reminders without an explicit warning that they create ongoing automated actions. Users may not realize reminders will continue firing until changed or removed, which can create unwanted persistence, leakage through notifications, or operational surprise.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Low
Category
Not specified by scanner
Confidence
77% confidence
Finding

The manifest describes a lifecycle that includes persistent storage and ongoing review, so saving plans is expected. However, this CLI goes beyond just generating and reviewing plans by interactively persisting goals, alignments, and completion state directly to local files, which is not explicitly conveyed in the command descriptions shown in the module documentation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

update_task_status reads and rewrites the target Markdown file in place, changing task completion state, but the method provides no user-facing notice, confirmation, or warning beyond an internal docstring. This is a filesystem-modifying operation that affects user data and lacks visible disclosure in the code.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

add_alignment edits the specified Markdown file by inserting or appending alignment records, but there is no user-facing warning, prompt, or logging indicating that the file will be changed. This is a persistent write to user data and should be disclosed more clearly.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
77% confidence
Finding

This code file contains a natural-language string assertion for "原始目标 / Raw Goal", indicating the underlying skill output may impose a bilingual locale format. Under the policy rule for language/locale, there is no evidence in this file of user opt-in or a documented justification for the forced locale choice.

Content

No source excerpt is available for this finding.