Back to skill

Security audit

Isfj Coach

Security checks for vulnerabilities and agentic risk

Overview

This non-executable coaching skill is not malicious, but it has serious packaging mismatches and tells the agent to keep sensitive coaching records without clear consent or deletion controls.

Review this carefully before installing. The artifact does not contain executable malware, but users may install the wrong package if they follow the README, and the skill should not save coaching records or build a profile unless the user explicitly agrees and can inspect or delete it. The publisher should align the ISFJ/INTP identity across all files and fix the install command before release.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:323
Finding

Persistent Retention of Sensitive Coaching Data Without Defined Safeguards

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 323-335
Vulnerability Type: Uncontrolled retention of potentially sensitive user information
Risk Level: Medium

Vulnerable Code Snippet

The following is an English translation of the complete relevant source segment:

markdown
### Feature 3: Growth Records (User Profile)

**Recorded content**:
- User's core problems (confusion, blockers, goals)
- User's areas of interest (areas where the user actually spends the most time)
- User's action records (what the user said they would do and what they actually did)
- User's key insights ("aha moments" from conversations)

**Usage scenarios**:
- Review previous progress during the next conversation
- Discover user patterns, such as "the user always selects option 1 but does not execute it"
- Adjust strategies, such as "the user needs stronger external supervision"

Related retention instructions also appear at SKILL.md:431-458, including archiving the user's problem, insights, action plan, and execution results. README.md:50 describes saving conversations and actions as an automatically executed feature.

Technical Analysis

The Skill directs the Agent to create and reuse a persistent user profile containing behavioral, career, relationship, emotional, and potentially mental-health-related information. It does not specify:

  • Whether explicit user consent is required before storage
  • Which storage or memory system receives the information
  • How access to the retained information is controlled
  • How long records are retained
  • How users can inspect, correct, export, or delete records
  • Which sensitive fields must be excluded or minimized
  • Whether the data may be reused outside the original coaching context

This is an insecure data-lifecycle design. Although the Skill does not contain code that directly writes files or sends records externally, a capable Agent may interpret th ...[truncated 1376 chars]

Remediation
View remediation

Remediation Suggestions

  1. Make persistent recording strictly opt-in and ask for explicit consent before storing any information.
  2. Clearly identify the storage system, intended purpose, retention period, and parties or components that can access the data.
  3. Store only the minimum information required. Prefer short, user-approved summaries over full conversations or detailed personal profiles.
  4. Exclude sensitive health, relationship, identity, financial, credential, and other high-risk data by default.
  5. Provide commands or documented procedures to inspect, correct, and permanently delete stored records.
  6. Apply encryption at rest, access controls, tenant isolation, and audit logging in any supporting memory implementation.
  7. Require renewed consent before reusing stored information for a materially different purpose.
  8. Replace statements that imply automatic storage with language such as: “With your permission, I can save a brief summary for future coaching sessions.”

T08 · Insecure Dependencies

Warning
Location
README.md:18
Finding

Installation Documentation References an Unrelated Package

Content
View full analysis

Vulnerability Details

File Location: README.md, lines 18-23
Vulnerability Type: Package identity mismatch and dependency confusion risk
Risk Level: Medium

Vulnerable Code Snippet

The following is an English translation of the complete relevant source segment; the installation command is reproduced verbatim:

markdown
### Installation

```bash
clawhub install intp-coach
text

The audited package is identified differently in the project metadata:

```json
{
  "name": "isfj-coach",
  "version": "1.0.0"
}

SKILL.md:2 also identifies the Skill as isfj-coach.

Technical Analysis

The documented installation command requests intp-coach, while the audited project declares its package name as isfj-coach. The requested intp-coach package is not included in the audited artifact, so its ownership, content, and integrity cannot be established from this project.

This creates a supply-chain trust-boundary failure: users reviewing the current artifact may reasonably believe the installation command installs the reviewed Skill, but the package manager will resolve a different registry identifier. If the other identifier is controlled by an unrelated or malicious publisher, its contents could execute with whatever capabilities the Agent platform grants installed Skills.

Additional package-integrity inconsistencies reinforce the risk: README.md and references/intp-insights.md describe an INTP coach, while the primary package metadata describes an ISFJ coach. The repository URL in clawhub.json:12 also appears unrelated to the declared package name.

Attack Path

  1. A user reviews or discovers the isfj-coach project.
  2. The user follows the installation instructions in README.md.
  3. The package manager resolves and installs intp-coach rather than the audited isfj-coach package.
  4. An unrelated publisher, or an attacker controlling that package identifier, suppli ...[truncated 904 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace the command with the correct package identifier:
bash
clawhub install isfj-coach
  1. Pin the installation to a reviewed version or immutable integrity identifier where the package manager supports it.
  2. Verify registry ownership of isfj-coach before publication and document the verified publisher.
  3. Align README.md, SKILL.md, clawhub.json, tags, keywords, and reference documents around one package identity and personality type.
  4. Replace the apparently unrelated repository URL with the canonical source repository for this package.
  5. Add a release check that rejects mismatches between the manifest package name and installation commands in documentation.
  6. Remove undeclared or nonexistent package entries such as the examples path in clawhub.json unless they are intentionally included.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (17)

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The README describes an INTP-focused coaching skill while the manifest declares an ISFJ coaching skill, creating a clear identity and scope mismatch. This can cause users, reviewers, or downstream systems to invoke the wrong behavioral guidance, weakening trust and potentially delivering inappropriate psychological advice to the wrong audience.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The documented use cases, examples, and coaching logic are explicitly tailored to INTP users rather than the declared ISFJ audience. In a coaching context, this is dangerous because personality-targeted guidance may be misapplied, leading to unsuitable advice, user confusion, and unsafe assumptions about user needs or emotional state.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The file content is entirely about INTP coaching, while the skill metadata states the skill is for ISFJ-specific guidance. This mismatch can cause the agent to deliver systematically incorrect persona-targeted advice, undermining reliability and potentially causing harmful guidance around boundaries, self-care, or change resistance because the recommendations are optimized for a different personality profile.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README states that the skill saves conversation and growth records, but it does not explain consent, storage location, retention period, deletion controls, or handling of potentially sensitive mental-health-adjacent data. Users may disclose intimate personal information during coaching, so undocumented persistence increases privacy and compliance risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill advertises a '成长记录' feature that saves dialogue and action history, but nowhere around the feature description does it clearly tell users that their data may be persisted, for how long, who can access it, or how they can opt out. Because this skill handles reflective coaching content that may include emotional state, habits, goals, and personal struggles, silent retention materially increases privacy and trust risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

These lines describe recording key insights into a user profile for later reuse without a clear consent step or warning. Profiling a user's recurring problems, interests, and behavior patterns in a coaching context can expose sensitive personal information if retained or surfaced unexpectedly in later sessions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

This section instructs the system to store core problems, interests, action records, and 'patterns' across sessions, which amounts to persistent behavioral profiling. The combination of sensitive self-improvement data and lack of explicit privacy warning or consent makes this a genuine privacy vulnerability, especially if users do not expect durable memory.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Persistent 'user profile' instructions create a data leakage risk because later responses may retrieve and reveal prior sensitive disclosures outside the user's immediate intent. In a coaching skill, stored notes about emotional struggles, goals, weak points, and execution failures are especially sensitive and can leak through summaries, references, or accidental cross-context reuse.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The end-of-conversation workflow mandates archiving the user's problem, insights, and action plan without notifying the user at the point of collection. End-of-session summarization is common, but mandatory hidden archiving of coaching disclosures increases the chance that sensitive information will be retained beyond the user's expectation.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Mandatory archival at conversation end increases the probability that personal disclosures are preserved and later exposed through unintended recall, mishandling, or overly broad memory use. Because the archived material includes user problems, insights, plans, and execution outcomes, a leak would reveal intimate behavioral and psychological context rather than trivial metadata.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The declared behavior in the manifest does not match the stated skill purpose, indicating the package may be mislabeled, repurposed, or incorrectly assembled. In an agent ecosystem, this creates a security-relevant integrity problem because selection, authorization, and user expectations may rely on metadata, leading to accidental invocation of an unintended coaching workflow.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The manifest metadata materially contradicts the declared purpose of the skill: it is presented as an ISFJ coaching skill, but the description, keywords, and tags reference INTP/general coaching content. This can mislead users and routing systems into invoking the wrong skill, causing inappropriate guidance and weakening trust and safety review because the manifest no longer reliably describes the artifact.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

This markdown file contains user-facing coaching guidance exclusively in Chinese from start to end, and there is no indication that the user can opt into another language or that the skill is intentionally limited to a Chinese-speaking context. The policy specifically calls out forced language or locale constraints as natural-language violations when no opt-in or justification is provided.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
79% confidence
Finding

This reference file is entirely written in Chinese and explicitly states it is background knowledge for the skill's internal use. For a general-purpose coaching skill, embedding a single-language operating context without any indication of user language choice can amount to a locale/language policy issue if the skill thereby defaults to Chinese behavior regardless of user preference.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
99% confidence
Finding

The reference file for an ISFJ coaching skill contains traits, functions, and coaching advice that align with a different personality profile, creating a clear specification-to-content mismatch. This can systematically misguide the agent’s behavior, causing it to give incorrect or psychologically inappropriate advice to users who requested ISFJ-specific guidance.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

Because the skill is explicitly marketed as ISFJ coaching, using mismatched reference guidance means the model may provide advice optimized for another personality pattern while presenting it as tailored support. In a self-help/coaching context, this undermines trust and can lead users to adopt unsuitable coping strategies, especially around boundaries, self-care, and change resistance.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The README and all example interactions are written entirely in Chinese, with no indication that users may choose another language or locale. Under the stated policy, forcing a specific language without user opt-in can be a natural-language policy violation unless the locale restriction is clearly documented and justified.

Content

No source excerpt is available for this finding.