Back to skill

Security audit

Gstack AI虚拟工程团队

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed engineering workflow, but it gives an agent broad automatic authority to edit, test, deploy, send code to other models, and persist lessons without enough control gates.

Review before installing. Use this only in repositories where automatic edits, tests, and release steps are acceptable. Disable external cross-model review unless you explicitly approve the provider and exact content being sent, keep production testing read-only, require confirmation before deployment or destructive requests, and do not allow automatic writes to global memory.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (4)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
tools/xcheck.md:24
Finding

Automatic Cross-Model Review Can Disclose Sensitive Project Content to External Providers

Content
View full analysis
Send to primary model: "Review this plan and identify issues" | | | +--> Primary model feedback | +--> Send to cross-check model: "Review the same plan independently" | +--> Cross-check model feedback ``` The cross-check prompt includes the reviewed material directly: ```markdown ## Content to Review [Paste code or plan here] ## Review Focus 1. Logic errors 2. Security issues 3. Performance risks 4. Maintainability ``` The example project configuration enables this behavior: ```yaml xcheck: true ``` ### Technical Analysis The cross-check workflow sends project plans or source code to an additional configured AI model. The instructions do not require: - Explicit user consent before transmission - Disclosure of the destination provider - A provider or region allowlist - Source-code sensitivity classification - Secret, credential, or personal-data scanning - Redaction or minimization of the transmitted payload - Confirmation that the secondary provider is covered by the project's confidentiality policy Because `xcheck` is enabled in the example configuration and is intended to run at critical decisions, after complex implementations, and before release, substantial proprietary material can cross the project's trust boundary. ### Attack Path 1. A project contains proprietary code, internal architecture details, customer-derived examples, or credentials embedded in configuration or source files. 2. The workflow reaches a configured cross-check point, such as a complex implementation or pre-release review. 3. T ...[truncated 901 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
roles/reflect.md:37
Finding

Project-Controlled Retrospective Content Can Poison Persistent Global Agent Memory

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
tools/freeze.md:30
Finding

Freeze Path Validation Can Be Bypassed Through Prefix Collisions and Symbolic Links

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
roles/test.md:23
Finding

Real-Data E2E Testing Allows Destructive Side Effects Without Mandatory Environment Isolation

Content
View full analysis
record reproduction -> automatically fix -> write regression test -> verify again -> pass ``` ### Technical Analysis The test role prioritizes real environments, real HTTP requests, and actual database mutation, but it does not mandate: - Exclusion of production environments - Disposable databases or isolated namespaces - Synthetic rather than customer data - Test-only accounts and API credentials - Transaction rollback - Cleanup of created or mutated records - Backups or restore validation - Suppression of real notifications, payments, or third-party calls - Explicit approval before destructive or state-changing requests - Rate and concurrency limits Because production is listed as a valid report environment and database deletions are explicitly included in side-effect verification, the workflow can cause real-world ch ...[truncated 1484 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (21)

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
70% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · tools/careful.md (reported line 14)May include surrounding context.

md
| 删除目录 | 数据永久丢失 | `rm -rf`, `shutil.rmtree`, `delete_file(目录)` |
| 清空数据库 | 所有数据丢失 | `DROP TABLE`, `TRUNCATE`, `DELETE *` |
| 覆盖重要配置 | 服务不可恢复 | 覆盖 `.env`, `nginx.conf`, `systemd service` |
| Force push | 覆盖团队提交历史 | `git push --force` |
| 生产环境变更 | 影响真实用户 | 连接生产DB、修改生产配置 |

### 🟠 高风险(警告 + 二次确认)

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
90% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · tools/careful.md (reported line 42)May include surrounding context.

md
┌─────────────────────────────────────┐
│  ⚠️  CAREFUL MODE                   │
│                                     │
│  操作: rm -rf /opt/data/cache       │
│  风险等级: 🔴 极度危险               │
│  影响: 永久删除所有缓存数据           │
│                                     │

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
85% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · tools/careful.md (reported line 42)May include surrounding context.

md
┌─────────────────────────────────────┐
│  ⚠️  CAREFUL MODE                   │
│                                     │
│  操作: rm -rf /opt/data/cache       │
│  风险等级: 🔴 极度危险               │
│  影响: 永久删除所有缓存数据           │
│                                     │

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
90% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · tools/careful.md (reported line 42)May include surrounding context.

md
┌─────────────────────────────────────┐
│  ⚠️  CAREFUL MODE                   │
│                                     │
│  操作: rm -rf /opt/data/cache       │
│  风险等级: 🔴 极度危险               │
│  影响: 永久删除所有缓存数据           │
│                                     │

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger list contains broad natural-language phrases such as 'QA测试', '安全审计', and '多角色协作开发' that can match ordinary developer requests and activate the skill unexpectedly. Because this skill can drive build, review, browser testing, and deployment workflows, accidental invocation increases the chance of unintended high-impact actions under the wrong context.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill advertises standalone commands for browser access, deployment, edit-freezing, and guarded execution, but the warning model is not prominent or enforceable at the point of use. In context, this is more dangerous because the same document promotes automatic multi-phase execution ending in deployment, so a user may trigger browser-based validation or shipping actions without a clear, unavoidable warning about external side effects.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill states that saying '用gstack模式做XXX' can automatically run the full 7-stage pipeline, while also allowing phases to be skipped based on the agent's own judgment. This creates ambiguous, overly broad activation for a workflow that includes testing in a real browser and shipping/deployment, making it possible for routine requests to escalate into autonomous system-affecting actions without sufficiently explicit user intent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

This markdown file presents the role instructions, workflow, and output format entirely in Chinese, with no indication that the user can choose another language. Under the policy for natural-language violations, forcing a specific language without user opt-in is reportable unless the locale constraint is explicitly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill content is entirely written in Chinese and prescribes Chinese-language review/output conventions without indicating that the assistant should adapt to the user's preferred language. This can cause misleading or unusable outputs for users operating in other languages, reducing review quality and potentially causing misunderstandings in design or security-related collaboration.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The file is written entirely in Chinese and frames the role/instructions only in that language, with no indication that users may choose another language or locale. Under the policy for natural-language violations, a skill that implicitly requires a specific language without opt-in should be flagged.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This markdown skill is written entirely in Chinese and presents the role, instructions, examples, and output format only in Chinese, which effectively enforces a specific language for the interaction. The file does not offer an opt-in language choice or explain that the skill is intentionally region-specific, so it conflicts with the language/locale policy criteria.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file title and all operational instructions are written in Chinese, and the skill defines its review role and output format in that language without any indication that the user may choose another language. This creates a natural-language locale constraint that appears mandatory rather than opt-in, which matches the policy's language/locale violation category.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The file title and all operational instructions are written in Chinese, and the prescribed output/report templates are also fixed in Chinese. This creates a locale/language constraint without any opt-in or alternative language path, which matches the policy category for forced language behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The markdown instructions are written entirely in Chinese and present the role behavior and output format exclusively in that language, without indicating that the user may choose another language or locale. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The entire role definition and output specification are written as mandatory Chinese instructions, with no indication that the user may choose another language. This creates a natural-language locale policy issue because the skill appears to require a specific language without opt-in or justification.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The workflow explicitly includes deployment and release stages, including production-oriented flows like '部署上线', 'Ship', and canary-style release behavior, but it does not require an explicit user confirmation or warning before impactful changes. In an agent skill, this can normalize or trigger production-facing actions without clearly signaling operational risk, increasing the chance of unintended deployments or unsafe changes reaching real environments.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The TEST failure flow describes automatic bug recording, automatic fixing, regression-test writing, and retry loops, which implies the skill may iteratively modify code and persist outputs without an explicit consent boundary. This is dangerous because repeated autonomous edits can introduce unintended changes, hide root causes, or create a self-reinforcing loop of risky modifications without the user's informed approval.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The file presents the skill entirely in Chinese, including the title and operational instructions, with no indication that the user can choose another language or that the skill is restricted to a Chinese-language environment. This is a natural-language locale policy issue because it imposes a specific language by default rather than offering choice or documenting a justified constraint.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The skill hard-codes use of 'other domestic models' for cross-checking without indicating user consent, provider choice, or language/locale considerations. In practice this can cause sensitive code, architecture, or business data to be routed to unintended third-party providers, creating avoidable data exposure and compliance risk during review workflows.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

The title and role framing are written in Chinese and set the review persona in that language context, but the file does not indicate that language is optional or user-selectable. This can violate a language/locale policy when a skill implicitly requires a specific language without opt-in.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
77% confidence
Finding

Natural-language policy violations include forcing a specific language without user opt-in. This file presents all operational instructions in Chinese and does not state that the skill is Chinese-only, region-specific, or that users may choose another language.

Content

No source excerpt is available for this finding.