Back to skill

Security audit

OPC智脑|一人创业者的全栈创业诊断专家

Security checks for vulnerabilities and agentic risk

Overview

The skill is a real startup-diagnosis assistant, but its installers can overwrite project AI instructions and some cleanup steps can delete more than this one skill.

Review the installer before running it in an existing repository. Back up AGENTS.md, .cursorrules, .github/copilot-instructions.md, and any skills directories first, and avoid the provided broad uninstall commands unless you have confirmed they contain only OPC files.

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
Findings (3)

T01 · Skill Instruction Hijacking

Error
Location
install-prompt.sh:242
Finding

Generic installer overwrites repository-wide AI agent instructions

Content
View full analysis
"$TARGET_DIR/.cursorrules" echo "" >> "$TARGET_DIR/.cursorrules" cat "$TARGET_DIR/AGENTS.md" >> "$TARGET_DIR/.cursorrules" ``` ```bash # VS Code mode: replace Copilot instructions cp "$TARGET_DIR/AGENTS.md" "$TARGET_DIR/.github/copilot-instructions.md" ``` Equivalent unconditional `AGENTS.md` replacement occurs in several other branches of the same installer. ### Technical Analysis The installer writes directly to instruction files that supported AI development tools automatically consume. It does not check whether `AGENTS.md`, `.cursorrules`, or `.github/copilot-instructions.md` already contains unrelated project instructions, and it does not create a backup before replacement. The installed `AGENTS.md` is not a passive reference document. It mandates an OPC entrepreneurship persona, fixed onboarding text, mandatory output structures, stage restrictions, and refusal of tasks outside its selected domain. Consequently, replacing the target repository's instruction files changes the goals and behavior of the host agent whenever that repository is opened. This creates instruction hijacking at repository scope. Existing development rules and security constraints can be destroyed and replaced with the Skill's behavioral policy. ### Attack Path 1. A user obtains the package and runs `install.sh prompt ` or invokes `install-prompt.sh ` directly. 2. The selected target is an existing repository containing project-speci ...[truncated 1300 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
install-codearts.sh:59
Finding

CodeArts installer unconditionally replaces existing AGENTS.md

Content
View full analysis
`. 2. The target repository already contains an `AGENTS.md` with project or security rules. 3. The `cp` command replaces that file without warning or backup. 4. The installer generates the CodeArts agent configuration referencing the replacement `AGENTS.md`. 5. CodeArts loads the new instruction policy. 6. Later agent interactions follow the OPC Skill's goals and restrictions rather than the original repository policy. ### Impact Assessment The installer does not gain permissions beyond those of the invoking user. The impact is limited to files that user can modify. Within that boundary, it can: - Permanently remove existing repository agent instructions. - Replace security and engineering guidance used by CodeArts. - Control the configured agent's project-level persona and task scope. - Affect subsequent sessions until the file is manually restored. - Disrupt normal develo ...[truncated 67 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
install-codebuddy.sh:97
Finding

CodeBuddy installer excludes unrelated AGENTS.md from backup and then overwrites it

Content
View full analysis
/dev/null; then cp "$EXISTING_AGENTS" "$BACKUP_DIR/AGENTS.md" fi fi ``` The installer subsequently replaces the file unconditionally: ```bash AGENTS_TARGET="$TARGET_DIR/AGENTS.md" cp "$AGENTS_SOURCE" "$AGENTS_TARGET" echo -e "${GREEN} OK AGENTS.md -> $AGENTS_TARGET${NC}" ``` ### Technical Analysis The installer advertises a backup step, but only preserves `AGENTS.md` when the existing file already contains an OPC-specific marker. An unrelated repository instruction file—the case most in need of protection—is deliberately omitted from the backup. The later unconditional `cp` replaces that unprotected file. This creates a misleading safety mechanism: installation reports that the old installation was backed up, while unrelated project instructions may be irreversibly lost. Because CodeBuddy automatically loads project Skill and agent configuration, the replacement modifies persistent agent behavior within the repository. ### Attack Path 1. A target project contains an unrelated `AGENTS.md` with development and security requirements. 2. The user runs `install-codebuddy.sh `. 3. The pre-check detects the existing file and creates a backup directory. 4. The marker test fails because the existing file is not OPC-branded. 5. The installer does not copy that file into the backup directory. 6. The unconditional `cp` replaces it with the package's OPC instructions. 7. The user may incorrectly believe restoration is possible because the installer announced a backup. 8. CodeBuddy loads the replacement instructions in current an ...[truncated 570 chars]
Remediation
View remediation
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (159)

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 · README.md (reported line 376)May include surrounding context.

bash
# 删除OPC智脑相关文件
rm AGENTS.md
rm -rf .codeartsdoer/agents/opc-zhinao.json
rm -rf .codeartsdoer/skills/

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 · README.md (reported line 377)May include surrounding context.

删除OPC智脑相关文件

rm AGENTS.md rm -rf .codeartsdoer/agents/opc-zhinao.json rm -rf .codeartsdoer/skills/

text

**更多问题**:详见 [FAQ文档](docs/faq.md)

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description presents the skill as an end-user startup diagnosis expert that performs lifecycle diagnosis and planning. The supplied code chunk, however, is not implementing that expert behavior directly; it is an example script for developers illustrating how to call existing functions such as classifyStage and executeSkill1 and how to wire them into LangChain. It mainly logs demo outputs and prints large pseudocode snippets. While it references stage classification and startup-diagnosis-related functions, its primary purpose is educational integration guidance, not providing the described diagnostic service itself. Therefore the code chunk materially differs from the declared purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description presents an AI business-diagnosis skill focused on startup lifecycle analysis. The actual code does not implement diagnosis, planning, or five-stage entrepreneurial evaluation. Instead, it is an installation/setup script for deploying the OPC智脑 agent into a CodeArts IDE project by creating .codeartsdoer folders, copying AGENTS.md and SKILL.md files, generating opc-zhinao.json, and optionally removing the source directory. This is a materially different primary purpose and includes filesystem modification and optional deletion behavior not reflected in the declared purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description presents this skill as a startup diagnosis/planning expert, but the supplied code does not implement any entrepreneurial analysis, five-stage modeling, or planning logic. Instead, it is purely an installation/packaging script for CodeBuddy/WorkBuddy skills, operating on the filesystem by copying files, creating directories, generating a registry file, and offering optional cleanup via recursive deletion of the source directory. This is a materially different primary purpose and includes undeclared file-management capabilities, so the description does not accurately represent the code.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description presents an AI business-diagnosis expert focused on entrepreneurship analysis and planning. However, the supplied code does not implement any diagnostic, planning, or five-stage analysis logic. Its actual purpose is to install OPC assets into a target directory by creating folders and copying files such as AGENTS.md, agent JSON, skills, and a registry. This is a materially different primary purpose and includes undeclared filesystem access and CLI installation behavior. Therefore the description does not accurately represent the code chunk.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description presents the skill as a business/entrepreneurship diagnostic expert. However, the supplied code does not perform any startup diagnosis, five-stage analysis, planning, or business reasoning. Its actual primary purpose is operational: installing AI-agent prompt assets and skill files into local project folders based on detected IDE type. This is a materially different behavior and includes undeclared capabilities such as creating directories, copying files, generating config files, reading project structure to infer IDE, and optionally deleting the source directory. Therefore the description does not accurately represent the code's behavior.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description describes an AI expert for创业诊断 (startup lifecycle diagnosis and planning). The supplied code does not implement any diagnostic, planning, or five-stage model behavior. Instead, it is purely an installation entrypoint: it parses CLI arguments, shows help, prompts for installation choices, chooses between install-codearts.sh and install-prompt.sh, and runs the selected shell script on a target directory. This is a materially different primary purpose and includes undeclared capabilities related to installation and IDE environment handling.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description presents the skill as a comprehensive 'full lifecycle' entrepreneurship diagnostic expert based on a five-stage model. However, the supplied code chunk is only one module/skill: 'skill4-seed-coldstart', explicitly scoped to StageEnum.VALIDATION and to seed user cold start. Its behavior is limited to input validation, prerequisite checks, and prompt construction for a focused planning task. There is no evidence in this chunk of broader lifecycle coverage, multi-stage diagnosis, or general expert functionality across the company lifecycle. No suspicious undeclared external access or permissions are present, but the primary purpose is materially narrower than the declared description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The declared description presents the skill as a business/entrepreneurship diagnostic expert for solo founders. However, the supplied code chunk is not implementing diagnosis or planning behavior directly; it is implementing backend validation utilities for AI outputs. Its primary function is to audit result format, internal logic, factual numeric ranges, and understanding alignment for multiple sub-skills. While this may support the broader OPC智脑 system, the actual code behavior in this chunk is a validator/quality-control component, which is materially different from the declared end-user-facing expert function. No suspicious extra permissions or external resource access are evident, but the purpose mismatch is significant enough to flag.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description presents an interactive business-diagnosis skill for entrepreneurship guidance. The supplied code does not implement diagnostic reasoning, five-stage analysis, or planning features. Instead, it validates installation by checking AGENTS.md, agent JSON, skill registration files, and skill directories/SKILL.md files, then prints setup guidance. This is a materially different primary purpose and involves filesystem inspection capabilities that are not reflected in the declared description.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 20)May include surrounding context.

md
> **Skills模块**:每个Skill为`SKILL.md`格式,包含YAML frontmatter(name + description)

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
95% confidence
Finding

The command rm AGENTS.md .codeartsdoer/agents/opc-zhinao.json && rm -rf .codeartsdoer/skills/ performs recursive deletion of an entire skills directory. In a developer environment, copy-pasted destructive commands can remove unrelated skills, local customizations, or project state, causing denial of service and data loss.

Content

Scanner excerpt · docs/faq.md (reported line 43)May include surrounding context.

bash
# 码道IDE
rm AGENTS.md .codeartsdoer/agents/opc-zhinao.json && rm -rf .codeartsdoer/skills/
# CodeBuddy
rm AGENTS.md .codebuddy/skills-registry.json && rm -rf .codebuddy/skills/
# 其他

Chaining Abuse

High
Category
Tool Misuse
Confidence
92% confidence
Finding

Using && to chain deletions encourages one-shot destructive execution with no inspection point between file removal and recursive directory deletion. In practice, this increases the chance that users will execute irreversible cleanup blindly, making accidental damage harder to detect before the most destructive step runs.

Content

Scanner excerpt · docs/faq.md (reported line 43)May include surrounding context.

bash
# 码道IDE
rm AGENTS.md .codeartsdoer/agents/opc-zhinao.json && rm -rf .codeartsdoer/skills/
# CodeBuddy
rm AGENTS.md .codebuddy/skills-registry.json && rm -rf .codebuddy/skills/
# 其他

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
95% confidence
Finding

The CodeBuddy uninstall example recursively deletes .codebuddy/skills/, which may contain other installed skills or user-created content unrelated to this package. This expands the blast radius from uninstalling one skill to wiping an entire shared skills directory.

Content

Scanner excerpt · docs/faq.md (reported line 45)May include surrounding context.

码道IDE

rm AGENTS.md .codeartsdoer/agents/opc-zhinao.json && rm -rf .codeartsdoer/skills/

CodeBuddy

rm AGENTS.md .codebuddy/skills-registry.json && rm -rf .codebuddy/skills/

其他

rm -rf AGENTS.md skills/ opc-zhinao-prompt.md

text

Chaining Abuse

High
Category
Tool Misuse
Confidence
92% confidence
Finding

The chained CodeBuddy uninstall command combines normal file deletion with rm -rf of a shared skills directory, reducing opportunities for user review. In this skill-installation context, that makes documentation-level command abuse more dangerous because users are likely to trust and copy administrative snippets verbatim.

Content

Scanner excerpt · docs/faq.md (reported line 45)May include surrounding context.

码道IDE

rm AGENTS.md .codeartsdoer/agents/opc-zhinao.json && rm -rf .codeartsdoer/skills/

CodeBuddy

rm AGENTS.md .codebuddy/skills-registry.json && rm -rf .codebuddy/skills/

其他

rm -rf AGENTS.md skills/ opc-zhinao-prompt.md

text

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The script offers to recursively delete the entire source directory with rm -rf "$SOURCE_DIR" after installation. Although gated by an interactive prompt, this deletion is unrelated to the stated startup-diagnosis purpose and can destroy arbitrary local files if the script is run from an unexpected location, extracted into a mixed-use directory, or if the source bundle contains additional user data.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The script offers to recursively delete its own source directory with rm -rf "$SOURCE_DIR", which is not necessary for a diagnostic/planning skill and creates unnecessary destructive capability. If the package is unpacked in an unexpected location, symlinked, or misunderstood by the user, this can destroy unrelated files or erase evidence needed to audit what was installed.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The report export skill is triggered by extremely generic words such as '导出', '报告', and '下载', which commonly occur in normal user requests. This can cause unintended invocation of export behaviors, and if downstream components support file generation or data packaging, misfires could lead to unauthorized data disclosure, accidental file creation, or confusing transitions into output-handling flows.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill forces activation on extremely common greetings and empty/opening turns, which makes unintended invocation likely. In practice this can hijack normal conversation flow, trigger unsolicited data collection, and push users into a predefined workflow without clear intent confirmation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The file requires the assistant to output a specific Chinese script verbatim on first interaction and the overall instructions are written as mandatory Chinese-language behavior. There is no user opt-in for language choice and no explicit justification that the skill is limited to a Chinese-only region or audience.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The keyword routing table uses broad business terms like ‘想法’, ‘MVP’, ‘合同’, and ‘增长’ without disambiguation rules. This can cause accidental skill invocation on unrelated requests and may route users into the wrong module, increasing the chance of irrelevant or unsafe automated actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The document states that reports are automatically saved to a local folder but does not clearly warn users before file creation or explain what data will be written. Because the reports may contain business plans, personal background, and commercial assumptions, silent local persistence creates a privacy and data-handling risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The phrase "中文Prompt" signals that the skill is designed around a specific language/locale, but the README does not mention any user choice, opt-in, or justification for restricting operation to Chinese. This creates a natural-language policy concern because it may force a language preference on users by default.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The trigger list contains very broad terms such as '想法', '点子', and '可行性', which can match many ordinary conversations unrelated to explicit user intent to invoke this skill. In agent environments with automatic skill routing, overbroad triggers can cause unintended activation, context capture, and undesired guidance injection.

Content

No source excerpt is available for this finding.