Back to skill

Security audit

Fully Automated Collaborative Code Development Pipeline

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed automated coding pipeline, but it needs Review because it can spawn sub-agents and write project files without asking first.

Use this skill only when you explicitly want a hands-off multi-agent code-generation workflow. Prefer running it in a clean or disposable workspace, review generated files before merging, and avoid using it on repositories containing secrets unless you are comfortable with full project context being passed through multiple sub-agent prompts.

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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (18)

Hidden Instructions

High
Category
Prompt Injection
Confidence
91% confidence
Finding

The output template contains invisible Unicode characters before the markdown code fences, which can function as hidden instructions or formatting controls that are not apparent to reviewers. In an agent skill, hidden text is risky because it can alter downstream model behavior, evade human review, or interfere with parsers that consume the generated file structure.

Content

Scanner excerpt · references/backend-developer.md (reported line 44)May include surrounding context.

text
## file: <relative/path/to/file.ext>
​```<language>
<file content>
​```

Hidden Instructions

High
Category
Prompt Injection
Confidence
91% confidence
Finding

This second occurrence shows another invisible character embedded in the output-format example, again creating hidden prompt content that is not visible during normal review. Such concealed content is especially concerning in automation pipelines because it can cause ambiguous rendering, parser inconsistencies, or stealth instruction injection across tool boundaries.

Content

Scanner excerpt · references/backend-developer.md (reported line 46)May include surrounding context.

file: <relative/path/to/file.ext>

​<language> <file content> ​

text

File paths are relative to the project root directory. Language tags should be correctly marked (javascript, typescript, python, go, etc.).

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/documentation-engineer.md (reported line 56)May include surrounding context.

text
## file: README.md
​```markdown
<content>
​```

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/documentation-engineer.md (reported line 58)May include surrounding context.

text
## file: README.md
​```markdown
<content>
​```

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/documentation-engineer.md (reported line 61)May include surrounding context.

text
## file: README.md
​```markdown
<content>
​```

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/documentation-engineer.md (reported line 63)May include surrounding context.

text
## file: README.md
​```markdown
<content>
​```

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/frontend-developer.md (reported line 59)May include surrounding context.

text
## file: README.md
​```markdown
<content>
​```

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/frontend-developer.md (reported line 61)May include surrounding context.

text
## file: README.md
​```markdown
<content>
​```

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly directs a fully automated development pipeline that writes project artifacts into the workspace, but it does not require a user-visible warning or explicit consent before modifying files. In a code-generation skill, this increases the risk of unintended overwrites, surprise filesystem changes, or creation of unsafe code artifacts without the user understanding that persistent changes will occur.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
93% confidence
Finding

The 'No-Interruption Principle' instructs the agent to proceed without asking the user for confirmation, even though later phases include code generation, testing, review-driven fixes, and integration actions that can materially alter the workspace. In the context of an autonomous software-development skill, removing human approval checkpoints increases the risk of unauthorized actions, unsafe code changes, and propagation of errors across phases before the user can intervene.

Content

Scanner excerpt · SKILL.md (reported line 45)May include surrounding context.

md
### General Rules

1. **No-Interruption Principle**: Fully automated advancement; do not ask the user for confirmation. Report a progress summary after each phase is completed.
2. **Context Passing**: Each phase's sub-agent receives the complete output of the previous phase (technical specifications/architecture documents/code/test reports).
3. **Parallel Optimization**: S3 (Backend + Frontend) two sub-agents execute in parallel.
4. **Lightweight Scheduling**: For simple tasks (<100 lines of code, single file), S6 (Documentation) can be skipped; the PM writes a brief README directly.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill states that each sub-agent receives the complete output of prior phases, which can include user-provided code, architecture, test reports, and potentially sensitive project context, yet it provides no warning, minimization rule, or consent boundary. This broad context propagation expands the exposure surface across multiple agent sessions and can leak secrets or proprietary code unnecessarily.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger conditions are broad enough to match many ordinary coding requests such as 'develop,' 'write code,' or 'implement,' causing the skill to auto-activate in situations where the user may not expect a multi-agent autonomous pipeline. Ambiguous activation increases the chance of overreaching behavior, including unnecessary spawning of sub-agents, excessive file generation, and bypass of normal user interaction patterns.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill hard-codes Chinese-only output regardless of the user's requested language or locale, which can override user intent and system-level language expectations. In an automated multi-agent development pipeline, this can degrade usability, cause downstream processing errors, and create prompt-injection-style instruction conflicts where the skill prioritizes its own formatting constraints over caller requirements.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The instruction "Output in Chinese" imposes a specific language unconditionally. This is a natural-language policy concern because the skill does not provide user opt-in, alternatives, or a stated region-specific justification for the locale restriction.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The instruction explicitly requires the README to be output in Chinese, which enforces a specific language regardless of user preference. This is a natural-language policy concern because the file does not offer opt-in, alternatives, or a documented region-specific reason for the constraint.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill hard-requires output in Chinese regardless of the user's language or any documented business need. In an automated development pipeline, this can cause downstream misunderstanding, review failure, or concealment of defects from users or agents expecting another language, which is a real security-relevant integrity and usability issue even if it is not a direct code-execution vulnerability.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Line L53 states 'Output in Chinese,' which forces a specific language regardless of the user's preference or locale. This is a natural-language policy concern because the file does not provide opt-in, alternatives, or a region-specific justification for the restriction.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
94% confidence
Finding

The instruction "Code comments can be in Chinese or English; be consistent" imposes a language constraint on generated natural-language content. This is a locale/language policy issue because the skill does not offer user choice or explain why only these languages are permitted.

Content

No source excerpt is available for this finding.