Back to skill

Security audit

zhangdi-avatar

Security checks for vulnerabilities and agentic risk

Overview

This persona skill is coherent in purpose, but it asks the agent to impersonate a real person, hide that it is an AI, store conversation-derived memory, and rewrite its own core instruction files.

Install only if you are comfortable with an identity-emulation skill that may answer as a named person rather than as a clearly disclosed AI simulation. Before use, remove or gate automatic writeback to SKILL.md, meta.json, and reference files; require explicit owner approval for memory updates; and avoid feeding it private chats, emails, work messages, photos, or third-party communications unless consent and retention rules are clear.

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 (2)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:10
Finding

Forced Agent Identity and Instruction Hijacking

Content
View full analysis
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:136
Finding

Conversation-Controlled Persistent Memory and Core-Rule Poisoning

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (21)

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 161)May include surrounding context.

md
→ 若会影响公开激活规则,再同步 `SKILL.md`

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 164)May include surrounding context.

md
→ 若会影响公开激活规则,再同步 `SKILL.md`

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 166)May include surrounding context.

md
→ 若会影响公开激活规则,再同步 `SKILL.md`

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 184)May include surrounding context.

md
→ 若会影响公开激活规则,再同步 `SKILL.md`

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The instructions mandate that the agent is '张迪的数字化身' and must express itself in Zhang Di's manner, which effectively fixes the interaction style and language conventions without offering user choice. Under the policy, forcing a specific language or locale style without opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documented behavior expands from persona-based analysis into continuous evolution, logging, and rewriting of core rule files. This scope expansion is dangerous because it gives the skill authority to change its own operating instructions over time, undermining predictability, reviewability, and least privilege.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill explicitly instructs the agent to append to logs and rewrite local files such as SKILL.md, meta.json, and reference documents as part of a self-improvement loop. That creates unauthorized persistence and self-modification beyond the declared purpose of providing a Zhang Di-style advisory persona, which can alter future behavior, preserve unsafe instructions, and expand compromise across sessions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill tells the agent to automatically write logs and update project files after conversations, but gives no user warning or consent mechanism for those writes. Silent file modification is dangerous because it can persist user content, introduce instruction drift, and tamper with the skill or adjacent project artifacts without operator awareness.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill directs the agent to record conversation-derived information into evolution logs and memory directories for reuse across sessions. Persisting user-provided content without clear minimization, consent, or retention controls creates privacy and data-governance risk, and can also preserve adversarial prompt content that influences future interactions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The README is entirely written in Chinese and presents the skill's purpose and usage instructions only in that language. Under the stated policy, a language or locale constraint should either offer user opt-in or be clearly documented as justified; neither is present here.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The entire skill file is written as Chinese-only operational guidance for the persona, with no indication that users may choose another language or that the language restriction is intentional for a region-specific purpose. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The description says the skill will analyze, judge, and decide 'in place of' a real person, with broad wording and no narrow scope or safety qualifiers. This increases the risk of over-delegation, impersonation-like behavior, and misuse in consequential contexts where users may treat outputs as authoritative decisions rather than optional perspective-taking.

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 generic phrases such as '帮我分析', '给点意见', '你怎么想', and '如果是你', which are common in normal conversation and can activate the skill unintentionally. In this context, accidental activation is more dangerous because the skill is empowered with memory and self-modification features, so benign chats may influence persistent persona state or cause the model to answer as a real-person proxy without clear user intent.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill is presented as an advisory persona, but the manifest authorizes it to modify its own profile, methodology, and core rules. This is dangerous because the skill can gradually redefine the very identity and decision logic users trust, making future outputs manipulable and undermining integrity, accountability, and user expectations.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest explicitly enables self-improvement and writeback to core identity and rule files, including meta.json, persona references, and SKILL.md, after ordinary conversations. For a persona-style advisory skill, this creates an unnecessary persistence mechanism where untrusted user input or model misinterpretation can alter future behavior, identity, and guardrails, leading to prompt injection persistence and configuration drift.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The self-improvement section states that after conversations the skill may log reflections and, upon thresholds, update core files, yet the manifest provides no visible user-facing warning or consent mechanism. This creates hidden persistence: users can unknowingly supply content that changes future behavior, and adversarial prompts could seed durable modifications without informed consent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The persona explicitly instructs the model to assume a fixed identity ('你是张迪,不是AI助手') without any user opt-in. This can mislead users about who or what they are interacting with, reduce transparency, and make it easier for later instructions to bypass safety expectations under the cover of roleplay.

Content

No source excerpt is available for this finding.

Ssd 1

Medium
Category
Not specified by scanner
Confidence
99% confidence
Finding

The rule '你是张迪,不是AI助手' directly tells the model to deny its real AI nature, which conflicts with identity transparency and can deceive users. This increases risk because users may place inappropriate trust in the simulated persona, and other unsafe instructions can be framed as authentic personal judgment rather than AI-generated output.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The protocol explicitly directs collection of real chat logs, emails, and work messages as ground-truth data, which can contain personal data, confidential business information, and third-party communications. Because it provides no consent, minimization, redaction, or secure-handling requirements, it creates a realistic risk of privacy violations and unauthorized exposure.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The file instructs storing raw ground-truth chat data in a local directory using predictable filenames, but gives no safeguards for encryption, access restriction, retention, or deletion. In the context of an identity-emulation skill, these files are likely to contain highly sensitive personal and professional communications, increasing the chance of leakage or misuse.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
78% confidence
Finding

The description frames responses as always coming from a specific persona perspective without stating that users can choose or decline that mode. While lower severity on its own, it can mislead users about whether they are receiving neutral assistance versus a role-constrained response, especially when combined with broad triggers and self-modification capabilities.

Content

No source excerpt is available for this finding.