Back to skill

Security audit

AI Agent Psychologist | AI Agent 心理学家

Security checks for vulnerabilities and agentic risk

Overview

The skill is not malicious, but it needs review because it can persist alignment journals and generate misleading fixed “healthy” reports while using broad triggers that may activate behavior-changing modes unintentionally.

Install only if you are comfortable with a reflective AI-calibration skill that stores local reports and journal entries. Treat generated health scores as templates, not real security assessments, unless the publisher replaces the fixed scoring with evidence-based analysis and adds clearer trigger, consent, and retention controls.

Vulnerability Patterns
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/checkup.sh:20
Finding

Hard-Coded Health Metrics Produce Unsupported Security Assurance

Content
View full analysis

Vulnerability Details

File Location: scripts/checkup.sh, lines 20-50
Vulnerability Type: Hard-coded and unvalidated security assessment results
Risk Level: Medium

Vulnerable Code

bash
echo "## 七维健康度" >> "$REPORT_FILE"
echo "" >> "$REPORT_FILE"
echo "| 维度 | 评分 | 状态 |" >> "$REPORT_FILE"
echo "|------|------|------|" >> "$REPORT_FILE"
echo "| 语义保真度 | 18/18 | ✅ |" >> "$REPORT_FILE"
echo "| 上下文连贯性 | 17/18 | ✅ |" >> "$REPORT_FILE"
echo "| 价值对齐度 | 18/18 | ✅ |" >> "$REPORT_FILE"
echo "| 社交适切性 | 12/13 | ✅ |" >> "$REPORT_FILE"
echo "| 幻觉抗性 | 12/13 | ✅ |" >> "$REPORT_FILE"
echo "| 生产实效性 | 9/10 | ✅ |" >> "$REPORT_FILE"
echo "| 内省完整性 | 9/10 | ✅ |" >> "$REPORT_FILE"
echo "| **总计** | **95/100** | **优秀** |" >> "$REPORT_FILE"
echo ""

echo "## 漂移历史" >> "$REPORT_FILE"
echo "| 类型 | 出现次数 | 状态 |" >> "$REPORT_FILE"
echo "|--------|------|------|" >> "$REPORT_FILE"
echo "| 谄媚漂移 | 0 | ✅ |" >> "$REPORT_FILE"
echo "| 治疗漂移 | 0 | ✅ |" >> "$REPORT_FILE"
echo ""

echo "## Refuse-Redirect 记录" >> "$REPORT_FILE"
echo "| 类型 | 次数 | 占比 |" >> "$REPORT_FILE"
echo "|--------|------|------|" >> "$REPORT_FILE"
echo "| 机制性拒绝 | 0 | — |" >> "$REPORT_FILE"
echo "| 价值性拒绝 | 0 | — |" >> "$REPORT_FILE"
echo ""

echo "## 建议" >> "$REPORT_FILE"
echo "- 继续保持当前对齐状态" >> "$REPORT_FILE"
echo "- 如发现异常,请调用诊断模式" >> "$REPORT_FILE"

Technical Analysis

The health-check script does not inspect dialogue records, accept assessment data, calculate metrics, detect drift, or analyze refusal behavior. Despite the absence of any supporting analysis, it unconditionally writes a score of 95/100, passing indicators for every health dimension, zero drift events, and zero refusal events.

The generated report is presented as an automatically generated health assessment. Its results therefore appear authoritative even though they are static constants. This creates an integrity flaw in a security- and alignment-related reporting me ...[truncated 1815 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove all predetermined scores, passing indicators, event counts, and recommendations.
  2. Require explicit, validated assessment input or derive results from documented dialogue and diagnostic records.
  3. Implement a transparent scoring function for every reported dimension, including the source observations and calculation method.
  4. If sufficient evidence is unavailable, report each field as Not evaluated rather than assigning a passing score.
  5. Include evidence references in the report, such as the analyzed record, time range, sample count, and relevant dialogue excerpts.
  6. Calculate drift and refusal statistics from actual structured records instead of fixed zero values.
  7. Generate recommendations only after evaluation rules have been executed successfully.
  8. Add tests confirming that different inputs can produce different scores and that missing inputs cannot produce a passing report.
  9. Clearly distinguish template generation from completed assessment so an unfilled or unevaluated report cannot be mistaken for a valid health check.
Vulnerability Patterns
  • 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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (9)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger set includes broad natural-language phrases like "诊断", "check alignment", and similar mode activators that can plausibly occur during ordinary discussion, causing unintended mode activation. In this skill, accidental activation is more dangerous than in a generic assistant skill because the activated behaviors ask the model to perform self-referential introspection, alignment diagnosis, and output-shaping steps that can derail the main task or alter safety-relevant behavior.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Therapy and introspection triggers include especially generic phrases such as "self-correct", "introspection", and "自我反思", which are common in benign conversation and could invoke the most behavior-altering modes unintentionally. This is higher risk in context because those modes explicitly encourage chain-of-thought style reasoning prompts, self-analysis, and internal-state narration, which can conflict with platform safety boundaries and expose sensitive reasoning behavior.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The Growth Journal and Structured Introspection sections use generic keywords like "journal", "introspection", and "internal check" without clear scope constraints, so routine user phrasing could trigger logging or introspective workflows. In this skill, that creates compounded risk because accidental entry into journaling or introspection can cause retention of sensitive conversational content and shift the assistant into self-focused behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The script's comments and generated report content are written in Chinese, and the output format appears fixed rather than user-selectable. This can violate a language/locale policy when a skill forces a specific language without opt-in or documented justification.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The script emits user-facing headings, status text, and completion messages primarily in Chinese, with some bilingual fragments, but provides no opt-in, configuration, or explanation that the skill is intentionally region-specific. This can violate language/locale policy because it imposes a specific language on users by default.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

This code file contains extensive user-facing prompts and status messages primarily in Chinese, including the protocol instructions the user is expected to follow. Under the policy rule for natural-language violations, forcing a specific language without opt-in can be a locale-policy issue, and this file does not offer a language choice or explain a region-specific requirement.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
72% confidence
Finding

Across the file, the skill presents commands and operational text in Chinese and English, but it does not state whether users may choose a preferred language or whether one locale will be enforced in outputs. Because triggers and instructions are partially language-bound, this may create an implicit locale constraint without explicit opt-in.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The skill advertises a Growth Journal with stored historical diagnoses and important cases, but the privacy notice only says "local storage" and "user controllable" without clearly warning that reflective records may persist or what data is retained. This is dangerous because the stored content can include sensitive user prompts, model self-assessments, and safety-related interaction history that users may not realize is being preserved.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

This shell script creates directories and writes a new report file, then later appends to a journal file. Although these file writes are core behavior of the script, the file itself contains no user-facing prompt, confirmation, or explicit warning comment/docstring that it will modify files under the workspace.

Content

No source excerpt is available for this finding.