Back to skill

Security audit

MDT Consultation

Security checks for vulnerabilities and agentic risk

Overview

This skill is useful for multi-agent review, but it automatically saves sensitive reports and can change skill or agent instruction files without approval.

Review before installing. Use this only where automatic multi-agent sharing and local memory retention are acceptable. For confidential contracts, medical information, legal work, or business strategy, require explicit user consent before saving reports, disable automatic edits to SKILL.md and SOUL.md, and keep any retained summaries redacted, scoped, and deletable.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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 (2)

other

Warning
Location
SKILL.md:87
Finding

Unconditional Persistence of Potentially Sensitive Consultation Reports

Content
View full analysis
_.md` - Includes: participating departments, each department's conclusions, comprehensive assessment, and final score ``` ### Technical Analysis The Skill requires every generated consultation report to be written to persistent memory. Archival is unconditional: the instructions provide no user-consent requirement, data-classification check, redaction process, access-control policy, retention period, or deletion procedure. This is security-relevant because the declared use cases include contract reviews, medical-plan reviews, internal official documents, strategic decisions, and legal assessments. Reports generated for these use cases can contain personal data, medical information, contractual terms, trade secrets, privileged legal analysis, or other confidential material. Persisting the compiled report is not required to complete the immediate consultation request. The broad archival instruction therefore increases the lifetime and exposure of sensitive information beyond the original processing context. ### Attack Path 1. A user submits a confidential contract, medical plan, internal document, or strategic proposal for consultation. 2. Specialized agents reproduce sensitive details in their findings and conclusions. 3. The coordinator incorporates those details into the final MDT report. 4. The Skill automatically writes the report to `memory/mdt-reports/_.md`. 5. A later agent, session, tool, or user with access to the shared memory path retrieves or inadvertently discloses the retained information. No separate privilege-escalation mechanism is r ...[truncated 638 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
references/eval.md:66
Finding

Unapproved Persistent Modification of Skill and Agent Instructions

Content
View full analysis
Remediation
View remediation
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 (8)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases include broad everyday expressions such as '大家一起看', '全面评审', and '会诊一下', which can cause unintentional activation in normal conversation. In this skill, accidental triggering is more dangerous because activation may launch multiple agents and invoke persistence behaviors, amplifying cost, data sharing, and privacy exposure without deliberate user intent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill description, trigger phrases, workflow labels, and example invocation are entirely specified in Chinese, and no alternative language option or user opt-in is offered. This can constitute a language-policy issue if the skill implicitly requires a specific language rather than allowing user preference.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill instructs automatic archival of every MDT report into persistent memory files, which creates a data-retention channel beyond the core purpose of generating a consultation output. Because this skill is designed for sensitive materials such as contracts, strategic decisions, medical plans, and leadership documents, silent persistence can store confidential or regulated data without user awareness or consent, increasing privacy, compliance, and data-exposure risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill description does not warn users that consultation reports will be automatically archived to memory files. This omission prevents informed consent about data retention and is especially risky given the stated use cases involve potentially sensitive legal, medical, technical, and strategic documents.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill directs automatic recording of failure analysis to an evolution memory file, introducing persistent self-optimization behavior outside the immediate user task. Even if intended for quality improvement, these records may capture user inputs, sensitive task context, or model shortcomings and retain them without notice, creating unnecessary long-term data collection.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

This markdown file presents all headings, usage scenarios, tasks, and output formats exclusively in Chinese. Under the policy rule for language or locale constraints, forcing a single language without explicitly offering user choice can constitute a natural-language policy violation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill omits a user warning that failed evaluations and optimization notes may be written to memory. Hidden retention of evaluation artifacts can expose sensitive task details and undermines transparency, even if the stored information is intended for internal improvement.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
81% confidence
Finding

This markdown file contains user-facing instructions and criteria exclusively in Chinese, and nowhere indicates that Chinese is optional or that the skill is intentionally restricted to a Chinese-speaking context. Under the natural-language policy rule, forcing a specific language without user opt-in can be a locale-policy violation.

Content

No source excerpt is available for this finding.