Back to skill

Security audit

test

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent user analytics tool, but it requires broad handling and reproduction of sensitive user conversation data without clear privacy controls.

Review before installing. Use this only on user data you are authorized to analyze, prefer redacted or aggregated outputs, avoid including full raw query logs unless explicitly approved, and install dependencies in an isolated environment with pinned versions.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (1)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:446
Finding
Unpinned Third-Party Dependency Installation## Vulnerability Details **File Location**: `SKILL.md:446` **Vulnerability Type**: Unpinned third-party dependencies **Risk Level**: Medium ### Vulnerable Code ```bash pip install pandas openpyxl requests python-docx ``` ### Technical Analysis The documented installation command resolves four packages from pip without exact version constraints, integrity hashes, or a reviewed lockfile. Consequently, installations performed at different times can retrieve different artifacts whose behavior was not covered by this audit. If an upstream package or its distribution account is compromised, a malicious release could execute code during package installation or when the installed library is imported. The instruction also does not require an isolated virtual environment, so following it may modify the invoking user's global Python environment. This finding concerns mutable dependency resolution. The audit found no evidence that any of the four named packages are currently malicious. ### Attack Path 1. An attacker compromises an upstream package release or its publishing account. 2. The attacker publishes a malicious version accepted by unconstrained dependency resolution. 3. A user follows the documented `pip install` command. 4. pip downloads and installs the attacker-controlled release. 5. Malicious installation or import-time code executes with the privileges of the user running pip. 6. The code can access resources available to that user, potentially including the sensitive AIME spreadsheets processed by the Skill. ### Impact Assessment Successful exploitation could permit arbitrary code execution with the invoking user's privileges. The affected scope could include files, credentials, environment variables, and network resources accessible to that account. If installation is performed as an administrator, the impact may extend system-wide; the documentation does not require elevated privileges. The project contains only `SKILL.md`, and the referenced scr ...[truncated 122 chars]
Remediation
## Remediation Suggestions - Place reviewed, exact dependency versions in a requirements or lock file. - Generate and verify cryptographic hashes for every permitted distribution. - Install with hash enforcement, for example: ```bash python3 -m venv .venv . .venv/bin/activate python3 -m pip install --require-hashes -r requirements.txt ``` - Review transitive dependencies and include them in the locked dependency set. - Use a trusted, explicitly configured package index or an internally controlled package mirror. - Add automated dependency vulnerability and integrity scanning. - Avoid running pip with administrator or root privileges. - Include the referenced scripts in the audited package so their dependency use and data-handling behavior can be reviewed.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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 (7)

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The skill processes detailed user interaction logs and directs generation of reports containing personal and behavioral information, but it provides no warning or safeguard language about sensitive data handling. This omission increases the chance that operators will process, share, or retain personal data without appropriate caution, authorization, or minimization.

Context-Inappropriate Capability

Medium
Confidence
95% confidence
Finding
The skill explicitly requires inferring sensitive attributes and statuses such as gender, age, internal-user status, competitor status, and crawler status from user conversation logs. These inferences go beyond routine behavior analytics, can be inaccurate, and create profiling/privacy risks without any stated necessity, consent, minimization, or review controls.

Ssd 3

Medium
Confidence
94% confidence
Finding
The skill requires the final report to contain a complete list of user queries with timestamps and translated versions for non-Chinese text. This creates a direct disclosure channel for raw user-generated content and behavioral metadata, which can expose sensitive financial interests, personal details, or other confidential information far beyond what is needed for aggregate analysis.

Natural-Language Policy Violations

Medium
Confidence
97% confidence
Finding
Forcing translation of every non-Chinese query without user opt-in removes a meaningful privacy control and expands the audience that can understand sensitive user content. In this context, the transformation is not necessary for all reports and can materially increase exposure of trading interests, personal details, and free-form user text.

Ssd 3

Medium
Confidence
91% confidence
Finding
The methodology repeatedly instructs the model to translate all non-Chinese user questions for the product team, normalizing broad disclosure of raw user inputs as part of routine reporting. This increases privacy risk because it encourages unnecessary propagation of complete user text to broader internal audiences.

Ssd 3

Medium
Confidence
90% confidence
Finding
The checklist operationalizes unconditional inclusion and translation of all non-Chinese questions in the final report, making broad disclosure a required output condition. Because the source material is user interaction logs, this materially raises the risk of privacy leakage, over-sharing, and downstream misuse of raw conversational data.

Missing User Warnings

Low
Confidence
88% confidence
Finding
The skill mandates translating all non-Chinese user queries and including them in reports without warning that original user content will be transformed and reproduced. That increases disclosure risk because translation can broaden accessibility of sensitive content and make it easier to redistribute outside the original context.

Static analysis

No suspicious patterns detected.