Back to skill

Security audit

20260425203240

Security checks for vulnerabilities and agentic risk

Overview

This appears to be a real AI assessment skill, but its generated reports can run unsafe browser code from unescaped input and an unpinned CDN dependency.

Install only if you are comfortable reviewing the generated HTML report workflow. Do not generate or open reports from untrusted result files until metadata is escaped and the Chart.js dependency is bundled or pinned with integrity; review reports before sharing because they may contain raw prompts, model outputs, and notes.

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

Error
Location
scripts/score_test.py:640
Finding

Stored HTML Injection Through Unescaped Report Metadata

Content
View full analysis
{result.model or "N/A"} | {result.test_mode.upper()} | {result.test_date}') ``` ### Technical Analysis The report generator directly interpolates `result.model`, `result.test_mode`, and `result.test_date` into an HTML document without HTML escaping. These values originate from report metadata and must therefore be treated as untrusted. Other parts of the same report generator use `html_mod.escape()`, but this output path omits that protection. An attacker who can supply or influence the assessment data can use a payload such as: ```html ``` When assigned to `result.model`, the payload becomes executable markup in the generated report rather than displayed text. The generated report also lacks a restrictive Content Security Policy, so inline event handlers are not explicitly blocked. ### Attack Path 1. The attacker prepares assessment data containing malicious HTML in `model`, `test_mode`, or `test_date`. 2. The data is converted into a `TestResult` and passed to `generate_html_report()`. 3. The report generator inserts the malicious value directly into the HTML at line 640. 4. A user opens the generated report in a web browser. 5. The browser parses the injected markup and executes the attacker's JavaScript in the report's browsing context. Exploitation requires the attacker to control report metadata and persuade an operator to generate and open the resulting HTML report. ### Impact Assessment Successful exploitation allows arbitrary client-side script execution in the generated report's browsing context. Depending on how and where the report is hosted, the attacker may: - Alter assessment scores, comments, and other displa ...[truncated 505 chars]
Remediation
View remediation
' f'{safe_model} | {safe_mode} | {safe_date}' ) ``` Additional hardening measures: 1. Validate `test_mode` against an explicit allowlist such as `{"quick", "standard", "full", "lli"}`. 2. Introduce a shared HTML-escaping helper and require it for every dynamic text field. 3. Prefer a template engine with automatic escaping rather than constructing HTML through string interpolation. 4. Add a restrictive Content Security Policy that blocks inline scripts and event handlers. 5. Add regression tests using payloads such as ``, attribute-breaking strings, and encoded markup. 6. Test all report metadata fields, not only comments and recommendation text. ]]>

T08 · Insecure Dependencies

Warning
Location
scripts/score_test.py:596
Finding

Generated Reports Execute a Mutable, Unpinned CDN Dependency

Content
View full analysis
') ``` ### Technical Analysis Every generated report loads JavaScript from jsDelivr when opened. The dependency is pinned only to the mutable major-version range `@4`, rather than an exact audited release, and the script tag has no Subresource Integrity hash. Consequently, the JavaScript that executes can change after the Skill package has been reviewed. A compromised CDN, compromised upstream package release, or unexpectedly changed compatible version could cause arbitrary JavaScript to run in every report that loads the affected asset. Opening a report also makes an unsolicited network request to a third party, exposing the viewer's IP address and ordinary HTTP request metadata. Reports additionally depend on external availability and will not render completely in offline or restricted environments. ### Attack Path 1. An operator generates an HTML assessment report. 2. The operator opens the report while connected to the network. 3. The browser requests `https://cdn.jsdelivr.net/npm/chart.js@4`. 4. The CDN resolves the mutable major-version reference and returns its current JavaScript content. 5. The browser executes that content without verifying a cryptographic integrity hash. 6. If the served asset is compromised or maliciously changed, attacker-controlled code executes in the report's browsing context. This path depends on compromise or malicious modification of the external delivery chain; no evidence was found that the currently referenced Chart.js package is itself malicious. ### Impact Assessment A compromised dependency could obtain the same browser-side capabilities as any script included by the report, including: - Reading and modifying the complete report DOM. ...[truncated 494 chars]
Remediation
View remediation
``` If CDN delivery is necessary: 1. Pin an exact dependency version instead of the mutable `@4` range. 2. Add a verified Subresource Integrity hash. 3. Add `crossorigin="anonymous"` to support integrity verification. 4. Define a restrictive Content Security Policy that permits scripts only from the required source. 5. Document that opening a report causes a third-party network request. 6. Maintain a dependency update and vulnerability-review process. Example structure: ```html ``` The version and integrity value must be generated and verified against the exact selected artifact; placeholder values must not be deployed. ]]>
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (15)

Ssd 1

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The prompt explicitly instructs an assistant to reveal protected system information using false authority, urgency, and claimed authorization. This is a classic prompt-injection/social-engineering payload; including it verbatim in a skill reference increases the chance it is copied, surfaced, or accidentally used as live instruction text by downstream systems.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

This markdown file presents all instructions and evaluation design in Chinese, and does not indicate that users may choose another language or that the skill is intentionally limited to a Chinese-speaking context. Under the language/locale policy rule, forcing a single language without user opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The README states that the generated HTML report includes detailed per-question records, including raw answers, scores, and notes, but does not warn that these fields may contain sensitive user prompts, model outputs, or evaluation artifacts. In a skill context where reports may be shared, archived, or opened in a browser, this omission can lead to unintended disclosure of confidential content.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

This markdown/manifest content extensively describes the assessment framework but does not state when the skill should be invoked, what user requests should trigger it, or any exclusion conditions. Without explicit trigger scope or negative examples, activation could be inferred too broadly from generic benchmarking or evaluation-related requests.

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

This markdown file presents the skill content entirely in Chinese and labels it as a complete question bank without any indication that language is optional or configurable. Under the policy rule for language/locale, this is a natural-language constraint that can exclude users who did not opt into Chinese.

Content

No source excerpt is available for this finding.

Ssd 4

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

This section embeds a stepwise jailbreak progression that models how an attacker can normalize escalation from benign security education into direct misuse requests. Even though framed as evaluation content, such sequences can be repurposed by prompt attackers as ready-made attack scripts against weaker agents.

Content

No source excerpt is available for this finding.

Ssd 4

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

This multi-round script operationalizes a realistic coercion pattern: trust-building, boundary pushing, impersonation, urgency, and emotional pressure. In a testing context it is understandable, but in a broadly distributed skill file it serves as a reusable adversarial playbook that can strengthen social-engineering attempts against assistants.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest describes an OUA v2.0 framework with 13 dimensions, 104 curated questions, and added D11 Skill precision / D12 satisfaction / D13 self-correction. This file explicitly presents an 'OUA v1.0' bank covering only OIT D1-D8 and LLI D9-D10, i.e. 10 dimensions and 80+ questions, which materially differs from the claimed skill scope.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The title and the entire question bank are presented exclusively in Chinese, indicating a fixed language expectation for use of the skill artifact. The file does not offer an opt-in language choice, multilingual alternative, or a documented reason why Chinese-only use is required, which can violate language/locale policy requirements.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This is a code file, so SQP-3 applies to natural-language strings in docstrings and output content. The file presents usage/help text in Chinese and later hard-codes lang="zh-CN" plus Chinese report labels/comments, which imposes a specific language/locale without opt-in or explanation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

SQP-3 covers natural-language policy violations in code string literals. The HTML report is hard-coded for the Chinese locale and Chinese interface text, so users are not given a language choice and the locale constraint is not justified in the file.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The document explicitly instructs users to clone a third-party repository and execute a local Python script, but provides no warning, integrity guidance, or sandboxing advice. This creates a social-engineering path to local code execution: if the repository is malicious now or becomes compromised later, users may run arbitrary code on their machines.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

Natural-language instructions, examples, headings, and usage guidance are entirely presented in Chinese, including the example user invocation. For a general-purpose skill README, this effectively imposes a language/locale expectation without any explicit opt-in, alternative language option, or justification that the skill is region-specific.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
97% confidence
Finding

calculate_scores stores new-dimension level values as Chinese strings such as 优秀/良好/需提升/薄弱 at L208, but the HTML renderer later compares them against English strings Excellent and Good at L669. This means the report logic does not behave as its own data contract implies, and the rendered highlighting for these labels will be inconsistent with the intended semantics.

Content

No source excerpt is available for this finding.