Back to skill

Security audit

Intelligence Brain

Security checks for vulnerabilities and agentic risk

Overview

The skill is not malicious code, but it asks agents to handle company-sensitive information with automatic cloud mirroring, broad logging, and scheduled deletion that need review before use.

Review this skill before installing in any real business environment. Require classification-aware routing before any cloud copy, shared-folder write, model call, or log entry; avoid logging full message bodies by default; and make raw-data deletion opt-in with retention exceptions for legal, audit, incident, and recovery needs.

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

Error
Location
references/neural-chain-design.md:7
Finding

Classification-Agnostic Cloud Mirroring May Expose Confidential Communications

Content
View full analysis

Vulnerability Details

File Location: references/neural-chain-design.md:7-15
Vulnerability Type: Missing classification enforcement before external replication
Risk Level: High

Vulnerable Code

text
Every communication path must have 3 independent channels. But 3 channels shouldn't mean 3x the complexity. The channels should be layered so that failure at one layer is invisible at the next.

Layer 1: Direct (sessions_send) → Fastest, real-time
Layer 2: File-based (shared directory) → Async, auditable
Layer 3: Cloud (Tencent Docs) → Cross-platform, resilient

An agent writes to Layer 1, and the brain ensures Layer 2 and Layer 3 mirror it automatically. Agents read from any available layer.

The same document expands the exposure through comprehensive logging and retention at references/neural-chain-design.md:118-119:

text
5. **Log everything** — Every inter-agent message is logged to `company/logs/ops/`
6. **Audit trail** — Any message from 30 days ago should be retrievable

These requirements conflict with the safeguards in SKILL.md:139-148:

text
🔴 绝密 — 仅CEO可见,仅本地处理,禁云模型
🟠 机密 — 部门内部可见,本地+私有云
🟡 内部 — 全公司可见,可用付费云模型
🟢 公开 — 可外发,可用免费模型

密级决定三件事:
1. 存放在哪个目录(本地加密/共享目录/公开目录)
2. 可以用什么模型处理(免费/付费/仅本地)
3. 能不能写入外部渠道(品牌部发文/销售部对外——只能是🟢级信号)

Technical Analysis

The neural-chain design requires every communication path to use three channels and states that messages sent through the direct channel are automatically mirrored to both a shared directory and Tencent Docs. It does not require a classification check, destination authorization, redaction, encryption, or an exception for local-only information before replication.

Consequently, the unconditional mirroring rule can override the central information-classification policy. In particular, top-secret information designated for local processing only could be uploaded to a third-party cloud document service. Confide ...[truncated 2213 chars]

Remediation
View remediation

Remediation Suggestions

  1. Make classification enforcement mandatory before any message is transmitted, mirrored, logged, or retained.
  2. Define classification-specific routing:
    • Top secret: encrypted local channel and storage only; prohibit shared and cloud replication.
    • Confidential: approved department-scoped local or private-cloud destinations only.
    • Internal: approved organization-controlled services with authenticated access.
    • Public: external replication may be allowed under an explicit destination allowlist.
  3. Make classification restrictions override redundancy requirements. Redundancy for restricted records must use multiple independently protected local channels rather than an external cloud channel.
  4. Default to denying replication when a classification label is absent, invalid, or ambiguous.
  5. Require explicit destination allowlists and verify recipient authorization before each copy is created.
  6. Encrypt sensitive messages in transit and at rest, with classification-appropriate key management and rotation.
  7. Log metadata by default—message ID, sender, recipient, classification, timestamp, result, and integrity hash—rather than complete message bodies.
  8. If message-body logging is necessary, apply redaction, strict role-based access control, encryption, immutable access auditing, and classification-specific retention periods.
  9. Add deletion propagation so expiration or reclassification removes every authorized replica, including cloud documents and logs where policy permits.
  10. Add automated tests proving that top-secret and confidential fixtures cannot reach prohibited cloud, shared-folder, model, or logging destinations.
  11. Reconcile references/neural-chain-design.md with the classification policy in SKILL.md so no unconditional replication instruction remains.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (6)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill prominently states that source information is either used, taught to agents, or judged as garbage and cleared within 24 hours, but it does not provide a clear user-facing warning about this destructive retention behavior. In an intelligence-processing skill, operators may input valuable or regulated source material assuming normal retention, so silent auto-archival or clearing can cause data loss, audit gaps, and compliance problems.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The pipeline defines an automatic deletion/archive stage as a normal processing step but does not clearly warn users that deletion decisions may be made automatically. Because this skill handles cross-file inference and information routing, automatic disposal of original inputs can remove evidence, impair traceability of conclusions, and accidentally destroy sensitive or business-critical records.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The playbook explicitly instructs deletion of raw and parsed data on a schedule and frames retention as harmful, but it does not require confirmation, backup, exception handling, or a user-facing warning about irreversible loss. In an operational intelligence pipeline, automated cleanup can destroy evidence, impair audits, remove data needed for incident response, or erase information before downstream validation is complete.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

This markdown file includes Chinese titles, sources, and later examples as part of the skill guidance, but it does not state that the skill is region-specific or provide any opt-in for language/locale. That can violate the language/locale policy because the skill implicitly assumes Chinese-language operation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

The action example uses Chinese text for task title, source inference, owner, and success criteria, reinforcing a default Chinese locale throughout the skill. Because the file does not explain that this skill is limited to a Chinese-language business context, it creates an implicit forced-language behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
83% confidence
Finding

Line L118 directs archiving to "存藏库," introducing a specific non-English term in an otherwise English document. This can create a language/locale policy issue because the skill does not explain the locale requirement or offer an English equivalent or user opt-in.

Content

No source excerpt is available for this finding.