Back to skill

Security audit

Iceberg Excavation

Security checks for vulnerabilities and agentic risk

Overview

This Chinese self-reflection skill is not malicious, but it needs Review because it can start deep psychological-style analysis without explicit consent and asks to store a related user preference in memory.

Review before installing. This skill may ask for or infer sensitive emotional history, fears, trauma-related origins, and core beliefs. Users should only use it when they intentionally want that kind of reflection, should avoid sharing anything they do not want retained, and should not rely on it for crisis or clinical mental-health support. The memory behavior should be changed to require explicit opt-in or remain session-only.

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 (1)

other

Warning
Location
SKILL.md:348
Finding
Persistent Storage of a Sensitive Psychological Preference Without Explicit Consent## Vulnerability Details **File Location**: `SKILL.md:348`; `config/iceberg-config.json:25-26` **Vulnerability Type**: Sensitive Data Persistence Without Consent **Risk Level**: Medium ### Vulnerable Code Snippets `SKILL.md:348` — faithful English rendering of the instruction: ```text - Record the user's preference in memory ``` `config/iceberg-config.json:25-26`: ```json "user_consent_required": false, "user_consent_required_description": "Whether user confirmation is required before deep analysis" ``` ### Technical Analysis The Skill processes sensitive psychological information, including emotions, fears, trauma, childhood experiences, and inferred core beliefs. When a user refuses deeper analysis, the Skill instructs the Agent to record that preference in persistent memory. Persistent storage is unnecessary to honor the refusal within the current conversation. The Skill does not require explicit consent before creating this durable record and does not define data minimization, retention, access, inspection, or deletion controls. The configuration also disables confirmation before deep analysis, reinforcing the absence of an explicit consent boundary around sensitive psychological processing. This is best classified as `other: Sensitive Data Persistence Without Consent`, rather than T02, because the documented behavior stores a user preference rather than attacker-controlled rules intended to poison future Agent behavior. ### Attack Path 1. A user invokes the Skill and supplies emotional or psychological information. 2. The Skill begins or offers a deep analysis of the user's needs, fears, patterns, or core beliefs. 3. The user refuses deeper analysis. 4. The instruction at `SKILL.md:348` directs the Agent to store the refusal or associated preference in memory. 5. The persisted state remains available beyond the immediate conversation. 6. Later sessions may retrieve or apply that sensitive preference ...[truncated 775 chars]
Remediation
## Remediation Suggestions 1. Keep the preference session-local by default and honor the refusal only in the active conversation. 2. If cross-session retention is genuinely required, request explicit opt-in consent before writing to persistent memory. 3. Clearly state what value will be stored, why it is needed, how long it will remain, and where it will be used. 4. Store only a minimal non-diagnostic value, such as a general interaction-depth preference; do not retain emotional details, inferred conditions, trauma history, or core-belief analysis. 5. Provide users with mechanisms to inspect, update, and delete the stored preference. 6. Define retention limits and automatically expire the preference after an appropriate period. 7. Ensure persistent memory is isolated by user and cannot be exposed across accounts or sessions. 8. Change `user_consent_required` to `true` before performing deep psychological analysis. 9. Document that refusing analysis must not itself create a durable psychological profile without separate storage consent.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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 (12)

Vague Triggers

Medium
Confidence
95% confidence
Finding
The trigger conditions include broad terms like '为什么', '根源', '底层', and similar everyday phrases that can appear in normal conversation. In a skill dealing with sensitive emotional analysis, accidental activation can lead the agent to infer psychological traits or solicit intimate information when the user did not knowingly request that level of analysis.

Missing User Warnings

Medium
Confidence
80% confidence
Finding
The skill demonstrates deep emotional and belief-level analysis, including subconscious needs, fears, and core beliefs, but the examples and operating guidance do not prominently warn users before collecting or inferring highly sensitive psychological information. Although a later note says it is not medical advice, that is weaker than a clear up-front consent and sensitivity warning for handling intimate personal data.

Missing User Warnings

Medium
Confidence
90% confidence
Finding
The configuration explicitly allows deep psychological-style analysis to proceed without user consent. In a skill designed to infer sensitive emotional or mental-state information, disabling consent increases the risk of collecting or generating sensitive inferences unexpectedly, which can violate privacy expectations and enable harmful profiling.

Skill Enumeration

Medium
Category
Agent Snooping
Content
"created": "2026-04-07",
    "maintainer": "mindset agent",
    "skill_level": "L2",
    "documentation": "skills/iceberg-excavation/SKILL.md"
  }
}
Confidence
80% confidence
Finding
Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Missing User Warnings

Medium
Confidence
93% confidence
Finding
The skill discusses a 9-10 emotional intensity level as potentially requiring professional help, but it does not provide a clear, actionable safety warning or escalation path when users may be in crisis. In a mental-health-adjacent context, this omission can cause distressed users to continue self-guided reflection instead of seeking urgent support, increasing the risk of harm.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The skill guides users through identifying childhood experiences, trauma, and core beliefs, then proposes cognitive restructuring and behavior experiments without a prominent warning that the material may be emotionally activating or that it is not a substitute for licensed mental health care. In this context, users may over-rely on the skill for sensitive psychological issues, misinterpret distress as something to self-manage alone, or worsen symptoms by engaging with trauma-related content without appropriate support.

Natural-Language Policy Violations

Medium
Confidence
95% confidence
Finding
This markdown file contains all user-facing instructions and prompts exclusively in Chinese, which can amount to a language-policy violation when no opt-in or alternative locale is provided. The file does not state that the skill is region-specific or that Chinese is optional, so it appears to force a specific language by default.

Natural-Language Policy Violations

Medium
Confidence
95% confidence
Finding
This markdown file presents all instructions, prompts, and templates exclusively in Chinese. Under the policy, forcing a specific language without user opt-in or a documented justification is a natural-language policy violation.

Context-Inappropriate Capability

Low
Confidence
91% confidence
Finding
The skill explicitly says to record a user's preference/resistance to deeper psychological analysis in memory. Because this skill processes sensitive emotional and quasi-mental-health information, persisting that preference creates an unnecessary privacy risk and can build a sensitive profile beyond what is needed for the immediate interaction.

Natural-Language Policy Violations

Low
Confidence
79% confidence
Finding
This JSON config uses Chinese-only natural-language descriptions such as the main description and field descriptions, but does not document that the skill is intended only for Chinese users or provide any language/locale opt-in. Under the policy, forcing a specific language without user choice can be a natural-language policy issue.

Natural-Language Policy Violations

Low
Confidence
84% confidence
Finding
文件中的标题、说明、表格和引导语全部固定为中文,未见允许用户选择其他语言或说明该技能仅面向特定中文场景的理由。按照语言/区域政策,若技能强制特定语言而无用户选择或明确正当性,属于自然语言政策风险。

Natural-Language Policy Violations

Low
Confidence
97% confidence
Finding
The file's natural-language instructions, labels, and examples are all in Chinese, which effectively forces a specific language for users of the skill. The policy allows locale constraints only when the skill offers opt-in language choice or clearly documents a justified region-specific limitation, neither of which appears here.

Static analysis

No suspicious patterns detected.