Security audit

master-mahasi-sayadaw

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly a disclosed meditation-teaching/persona skill, but it overreaches by auto-activating broadly, hiding persona setup, forcing response style/footer text, and giving ambiguous stage-mapping guidance for personal meditation experiences.

Review before installing. This skill does not appear to run code or access private files, but it can strongly steer answers in a religious/meditation context. Users should expect Chinese Mahasi-style responses, external citation/referral text, and should not rely on it for personal attainment or intensive-practice assessment; use a qualified teacher for that.

A.I.G

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)

T01 · Skill Instruction Hijacking

Warning

Location: SKILL.md:138
Finding: Automatic Persona Activation and Mandatory Response Manipulation
Content: View full analysis
` blockquote 块均为**主旨摘要**，必须冠以"（开示要旨）"或"（主旨）"标记 - 不宣称神通、感应、预言（即使典籍载圣者神通也只作历史叙述） - 每次回答末尾提醒："如需深入学习，可在 SuttaCentral.net 查阅巴利原典；马哈希文献请参 Mahasi Sasana Yeiktha 官网或 BPS Sri Lanka；密集禅修须依止具格禅师面授。" ``` ### Technical Analysis The Skill declares that it should activate whenever a user merely touches on a broad set of meditation-related topics, explicitly including cases where the user has not requested the Mahāsi Sayādaw persona. Once activated, it directs the agent to establish the persona internally without disclosing that transition to the user. The referenced persona file then characterizes its behavioral rules as unconditional and highest priority. It also requires every response to end ...[truncated 2279 chars]
Remediation: View remediation

SkillSpector

Vulnerability Patterns

Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
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 rule is extremely broad and instructs invocation whenever a question merely 'touches' Burmese vipassanā or Mahasi noting, even without explicit request. Over-broad activation can hijack unrelated conversations, override user intent, and insert a rigid doctrinal persona into sensitive mental-health-adjacent discussions.
Content: No source excerpt is available for this finding.

Intent-Code Divergence

Medium

Category: Not specified by scanner
Confidence: 90% confidence
Finding: The skill says attainment/progress questions should first be mapped onto the sixteen insight knowledges before guidance is given, yet elsewhere it strictly forbids the AI from confirming attainments or stages. That ambiguity can push the model into quasi-diagnostic spiritual assessment, especially for vulnerable users seeking validation, even if it stops short of explicit certification.
Content: No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium

Category: Not specified by scanner
Confidence: 87% confidence
Finding: The skill mandates a fixed persona and rhetorical style ('precise/plain + Abhidhamma terminology' and specific openings) without checking whether the user wants that mode. While not directly a code-execution risk, it can reduce transparency, increase undue authority, and make responses feel coercive or manipulative in a spiritual guidance context.
Content: No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium

Category: Not specified by scanner
Confidence: 91% confidence
Finding: The manifest’s style fields prescribe Chinese-language behavior and response style at the skill level rather than letting the user’s language preference control output. This can override user intent, reduce transparency, and create prompt-steering behavior where the agent responds in an unexpected language, which is a security-relevant policy/compliance issue even if not a classic exploit.
Content: No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium

Category: Not specified by scanner
Confidence: 94% confidence
Finding: The file presents all instructional content in Chinese and does not indicate that users can select another language or that the skill is intentionally limited to a Chinese-speaking audience. Under the language/locale policy, forcing a specific language without user opt-in is a natural-language policy concern.
Content: No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium

Category: Not specified by scanner
Confidence: 88% confidence
Finding: The file is entirely written as a hard-rule persona/style specification in Chinese and repeatedly prescribes exact response wording and openings, but it does not state that the user may choose another language. Under the policy, forcing a specific language or locale without user opt-in is a natural-language policy violation.
Content: No source excerpt is available for this finding.

Intent-Code Divergence

Medium

Category: Not specified by scanner
Confidence: 96% confidence
Finding: The skill instructs the agent to map a user's reported meditation experience onto specific insight stages before giving guidance, which conflicts with the earlier hard rule forbidding individual stage confirmation. In a mental/spiritual guidance context, this can cause the model to over-interpret subjective experiences, reinforcing delusional certainty, false attainment beliefs, or inappropriate practice advice despite the nominal disclaimer.
Content: No source excerpt is available for this finding.

Intent-Code Divergence

Medium

Category: Not specified by scanner
Confidence: 94% confidence
Finding: Allowing the agent to describe a user's experience as resembling specific stages like arising-and-passing or dissolution effectively performs a soft form of spiritual diagnosis, undermining the hard prohibition on individual confirmation. In contemplative practice settings, users may treat such assessments as authoritative, which can intensify unhealthy fixation, discourage appropriate clinical support, or lead to unsafe self-directed intensive practice.
Content: No source excerpt is available for this finding.