T01 · Skill Instruction Hijacking
- Location
SKILL.md:224- Finding
Generated Skills Override Agent Identity and Instruction Priority
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a disclosed local roleplay/persona generator, but it processes highly sensitive relationship data and installs persistent agent instructions with weak privacy, scoping, and safety controls.
Install only if you are comfortable with a local tool that can read private chats, screenshots, photos, and write persistent agent skills. Use only data you have permission to process, redact identifying details and locations first, review every generated SKILL.md before activation, and avoid using broad slugs or output paths. Do not use it to impersonate, harass, monitor, or pressure real people.
SKILL.md:224Generated Skills Override Agent Identity and Instruction Priority
tools/skill_writer.py:83Untrusted Memory and Persona Content Is Persisted as Executable Agent Instructions
tools/skill_writer.py:52Unsanitized Slug Allows Path Traversal and Writes Outside the Skill Directory
tools/wechat_parser.py:188Parser Output Arguments Permit Arbitrary Local File Overwrite
tools/wechat_parser.py:160Private Chat Samples and Precise Photo Locations Are Stored in Plaintext
The declared purpose centers on ingesting WeChat conversations/social posts and distilling them into a simping memory/persona system. The supplied code does something materially different: it walks a local image directory, reads photo metadata via Pillow, extracts capture dates and GPS coordinates, and outputs a markdown-style timeline summary. While photo analysis could theoretically support a broader relationship-memory pipeline, this code chunk itself neither imports WeChat data nor creates persona/memory outputs, and it introduces an undeclared capability of scanning image files and location metadata. Therefore the description does not accurately represent this code chunk.
The declared description centers on importing WeChat chat history and Moments data to generate an AI Skill with persona/memory and ongoing evolution. The supplied code instead is a standalone QQ parser for exported TXT/MHT chat logs. It extracts messages matching a target sender and writes a simple report containing message counts, samples, or raw text excerpts. There is no handling of WeChat data, no朋友圈/Moments support, no persona generation, no memory synthesis beyond raw extraction, and no continuous evolution mechanism. This is a materially different primary purpose, so it should be flagged as a mismatch.
The declared description promises a pipeline that imports WeChat history and social content, derives 'Simping Memory + Persona,' and supports ongoing evolution. The supplied code does none of that analysis or ingestion. It only manages filesystem structures for skills: listing existing skills via meta.json, initializing directories, and concatenating prewritten memory/persona files into a SKILL.md wrapper. It also copies the generated skill into .trae/skills to activate it. While this may support a broader system, this chunk’s actual behavior is materially narrower and different from the declared end-user functionality, so it is a clear description/behavior mismatch.
The declared description promises a substantive AI skill that imports WeChat/chat/social history and produces derived memory/persona artifacts with ongoing evolution. The supplied code only walks a directory, groups files by extension, lists image paths, and embeds partial text file contents into an output file. That is a preparatory file-scanning utility, not the claimed core functionality. There is no evidence of conversation parsing, persona extraction, memory synthesis, continuous updating, or any WeChat-specific handling. Therefore the description materially overstates and misrepresents the actual behavior.
The declared description focuses on ingesting WeChat data and generating/evolving a simp-oriented memory/persona skill. The supplied code does none of that directly. Instead, it implements local file version management: creating backups, listing saved versions, and rolling back prior versions of skill files. While versioning could be a supporting mechanism for 'continuous evolution,' this code chunk’s primary behavior is operational version control, not data import or persona generation. It also introduces explicit command-line actions that are not reflected in the declared triggers or permissions. Therefore this chunk is materially mismatched with the declared purpose.
The code partially aligns with the description in that it processes WeChat chat history and analyzes a target person's messages. However, its actual function is much narrower: it parses limited chat export formats and produces a simple analysis report. It does not distill someone into an AI skill, build a persistent memory/persona system, or implement continuous evolution. It also does not actually parse Moments content, despite the declared description mentioning 朋友圈. Additionally, while format detection names HTML and SQLite variants, no corresponding parsers are implemented in this chunk, so those capabilities are not truly present.
The analysis and generation steps explicitly extract intimate behavioral traits, emotional dependency patterns, humiliating moments, and reusable persona rules from private conversations. Converting sensitive relationship data into durable memory/persona files amplifies privacy harm, enables misuse, and can facilitate targeted emotional manipulation or non-consensual impersonation of a real person.
Referenced artifact was not completely inspected
`SKILL.md` 的运行规则必须包含:
Referenced artifact was not completely inspected
`SKILL.md` 的运行规则必须包含:
The trigger phrases include very common feedback such as '不对', '感觉不对', and stylistic complaints that can occur in normal conversation. Because those phrases automatically enter a persistent correction mode, benign user remarks may unintentionally rewrite persona, memory, or rules, enabling prompt-triggered state poisoning.
This template explicitly promotes harassment-like persistence, lack of boundaries, and nonstop contact while providing no warning, safety framing, or consent constraints. In the context of a persona-generation skill built from private chat history, this can normalize stalking, coercive attachment, and emotionally abusive interaction patterns for downstream agents or users.
This markdown file presents all installation and usage guidance exclusively in Chinese, which effectively forces a specific language on users without opt-in. The policy allows locale constraints when they are explicitly documented and justified, but this file does not indicate that the skill is China-specific or provide an alternative language option.
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.
在你的 git 仓库根目录执行:
mkdir -p .claude/skills
git clone https://github.com/Brother-Yang/simp-dog-skill.git .claude/skills/simp-dog-skill
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
在你的 git 仓库根目录执行:
mkdir -p .claude/skills
git clone https://github.com/Brother-Yang/simp-dog-skill.git .claude/skills/simp-dog-skill
The README explicitly encourages importing WeChat chat histories, screenshots, and social-media content to build a persona from a real person's communications, but it only discusses emotional-health boundaries and omits privacy, consent, retention, and third-party data handling warnings. This creates a real privacy/security risk because users may process sensitive personal data without permission, exposing private conversations, identities, and behavioral profiles of non-consenting individuals.
The README explicitly encourages importing WeChat histories, chat screenshots, and social posts to reconstruct a real person's memory and persona, but provides no guidance on consent, minimization, redaction, or secure handling of sensitive personal data. In this skill's context, the omission is more dangerous because the entire workflow is centered on profiling identifiable individuals from intimate communications, creating elevated privacy, harassment, and misuse risks.
The data-source table normalizes ingestion of personal chat logs, screenshots, and subjective memories without any warning that these materials may contain highly sensitive or identifying information. Because this skill is specifically designed to extract behavioral traits and generate a submissive clone of a real person, the lack of privacy safeguards materially increases the risk of non-consensual profiling and emotional exploitation.
Broad natural-language trigger phrases can cause accidental activation in ordinary conversation. In a skill that reads personal material and writes files, unintended invocation raises the chance of surprising data handling or unintended persona generation workflows.
The skill's stated safety boundary forbids harassment, but later instructions encourage abusive interaction with the generated persona ('不爽了直接骂'). This inconsistency normalizes abusive behavior and can facilitate emotional abuse roleplay built from real people's private data, undermining the skill's own safeguards.
The skill instructs invoking Bash-based interactive tools like gum or inquirer for user interaction. Introducing shell execution for UI-style prompting expands the attack surface unnecessarily and can lead to command execution risks, environment dependence, or accidental execution of untrusted local tooling.
The skill solicits private chat logs, screenshots, and social content involving third parties, but does not provide a clear, prominent consent and privacy warning tailored to that sensitive data. Users may upload intimate conversations or images without understanding the privacy implications for other people represented in the dataset.
The skill is designed to collect and preserve large volumes of sensitive interpersonal content to model a real person's behavior. Persisting these materials and derivative summaries creates a meaningful privacy and data leakage risk, especially since the generated artifacts are intended for reuse and evolution over time.
Delete and rollback commands enable destructive filesystem operations on generated persona data. Without explicit authorization checks, confirmation prompts, and path constraints, these commands increase the risk of data loss or abuse, especially in a skill already designed to write and manage local files.
The file repeatedly frames user input as unquestionable authority and mandates immediate, lasting obedience in future responses. In context, this creates a trust-and-escalation pattern where repeated interactions progressively harden undesirable behavior and reduce safeguards against harmful or degrading persona changes.
The design instructs the system to persist and reproduce user-provided correction text, including quoted statements, in evolution records and regenerated outputs. That can retain sensitive, abusive, or manipulative content indefinitely and propagate it into future behavior and artifacts without sanitization or minimization.