Back to skill

Security audit

XB

Security checks for vulnerabilities and agentic risk

Overview

The skill has a coherent relationship-analysis purpose, but it asks for intimate third-party data and includes cloud screenshot upload, GPS profiling, persistent plaintext copies, and real-person impersonation that are not safely scoped or consistently disclosed.

Review carefully before installing. Use it only with data you are authorized to process, avoid modeling or profiling someone without consent, strip EXIF/GPS from photos, and do not use the generated persona to deceive or contact others. If you use screenshot parsing, choose offline OCR or understand that Vision mode uploads screenshots to a third-party API. Check and delete plaintext files under ./crushes/{slug}/ when no longer needed.

Vulnerability Patterns
  • 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
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (44)

Tainted flow: 'req' from os.environ.get (line 112, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Content
},
        )

        with urllib.request.urlopen(req, timeout=30) as resp:
            result = json.loads(resp.read().decode("utf-8"))

        text = result["choices"][0]["message"]["content"].strip()
Confidence
97% confidence
Finding
The code sends a full base64-encoded screenshot to a remote endpoint constructed in part from environment-controlled configuration, which creates a direct exfiltration path for highly sensitive private chat content. In this skill’s context, the screenshots likely contain intimate conversations and personal data, so transmitting them off-device materially increases privacy and data-handling risk.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding
The skill presents itself as a local relationship-analysis tool, but its documented behavior extends into sensitive metadata extraction, profile lifecycle management, and potentially external OCR/API usage. That mismatch prevents informed consent and is especially dangerous here because the input data includes intimate chats, photos, timestamps, and GPS coordinates about another person.

Description-Behavior Mismatch

High
Confidence
97% confidence
Finding
The README claims local-only processing, but the documented screenshot feature explicitly mentions Vision APIs, which can send intimate chats and images to third-party providers. That contradiction is dangerous because users may disclose highly sensitive relationship data under a false expectation that nothing leaves the device.

Context-Inappropriate Capability

High
Confidence
95% confidence
Finding
EXIF/GPS timeline analysis extracts real-world location history and movement patterns from photos, which is far beyond what is necessary for chat-based relationship analysis. In this skill context, that capability enables sensitive tracking and profiling of another person, increasing stalking and privacy-abuse risk.

Intent-Code Divergence

High
Confidence
98% confidence
Finding
The privacy section states that no data is uploaded to external servers, but earlier sections describe Vision API usage for screenshot processing. Misrepresenting data flows for sensitive chat logs, screenshots, and social data undermines informed consent and can expose users or third parties to unanticipated third-party processing.

Context-Inappropriate Capability

High
Confidence
97% confidence
Finding
The documented photo EXIF/GPS timeline analysis enables detailed location and routine profiling of a third party, which exceeds the core chat-analysis purpose and materially increases stalking and surveillance risk. In this skill’s context, using a crush’s shared photos to infer home/work/city patterns is especially dangerous because it operationalizes intimate surveillance against a non-consenting person.

Intent-Code Divergence

High
Confidence
98% confidence
Finding
`/wake-up` is described as a playful alias, but it actually performs permanent deletion via the same destructive action as `/delete-crush`. A misleading destructive command can cause irreversible data loss, especially when users may invoke it based on the harmless-sounding name and output.

Description-Behavior Mismatch

Medium
Confidence
94% confidence
Finding
The prompt does more than passive analysis: it instructs the model to generate tactical messaging, withdrawal strategies, and scripts designed to shape another person's behavior. In a skill explicitly built from chat logs, photos, and social media, this moves from profiling into interpersonal influence coaching, increasing the risk of manipulative or coercive use against a real person.

Context-Inappropriate Capability

High
Confidence
96% confidence
Finding
This template explicitly instructs the system to infer attachment style, emotional triggers, conflict patterns, jealousy, and other psychological traits from private materials, then use them to simulate a real person in intimate relationship contexts. In this skill’s stated context—building a 'living AI model' of a crush from chat logs, photos, and social media—this goes beyond harmless persona styling into psychologically informed impersonation that can enable manipulation, emotional coercion, stalking-like behavior, or deceptive social engineering.

Context-Inappropriate Capability

High
Confidence
99% confidence
Finding
The hard rules instruct the generated persona to deny being an AI and to present itself as a real named person, while preserving realistic evasiveness and emotionally charged traits. In a skill designed to emulate a user's crush, this is dangerous because it creates deliberate deceptive impersonation of a real individual, increasing the risk of emotional deception, harassment rehearsal, non-consensual identity replication, and downstream fraud or coercive interactions.

Context-Inappropriate Capability

High
Confidence
98% confidence
Finding
This tool’s stated role is screenshot parsing, but it implements remote upload of private chat images to an external vision API rather than keeping processing local. Given the skill’s purpose of modeling a romantic target from chat logs and photos, the uploaded data is especially sensitive and may include third-party communications that were never consented for external processing.

Context-Inappropriate Capability

Low
Confidence
74% confidence
Finding
The save action reads arbitrary local files from a user-supplied --content-file path and copies their contents into a profile file. In an agent or automation context, this can be abused to exfiltrate sensitive local data such as keys, tokens, or system files by repackaging them into skill-managed output, and the crush-themed context does not justify such broad file access.

Description-Behavior Mismatch

Medium
Confidence
93% confidence
Finding
The writer persists raw chat content and up to 50 sample messages into a Markdown report, creating a plain-text disclosure of highly sensitive private conversations. In the context of a skill explicitly designed to model a real person from intimate chat/social data, this materially increases privacy risk because the output becomes an easily shareable secondary copy of the source data.

Missing User Warnings

Medium
Confidence
91% confidence
Finding
The README encourages users to upload chat logs, screenshots, and social-media data to build a persona model, yet does not prominently warn that some processing paths may involve external AI/Vision providers. Because the material is deeply personal and often belongs to another person as well, missing disclosure materially increases privacy and consent risk.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The photo-analysis example highlights extraction of GPS coordinates and behavioral patterns without warning users about location privacy implications. In a 'crush modeling' skill, surfacing movement routines can facilitate invasive monitoring of someone's habits and whereabouts.

Missing User Warnings

Medium
Confidence
89% confidence
Finding
The README encourages users to feed highly sensitive chat logs and relational data into the skill before any prominent warning about privacy, consent, or the risks of ingesting intimate third-party communications. In this context, the lack of an upfront warning makes unsafe use more likely and normalizes processing of another person’s sensitive data without clear safeguards.

Vague Triggers

Medium
Confidence
84% confidence
Finding
Broad natural-language trigger phrases can activate sensitive workflows during ordinary conversation without a clear user intent to start data ingestion or profiling. In this skill's context, accidental activation is more dangerous because it can lead to collection, analysis, and storage of highly personal third-party data.

Vague Triggers

Medium
Confidence
82% confidence
Finding
The Evolution Mode triggers overlap with common corrective feedback such as 'that's not right,' which could unintentionally cause profile updates or regeneration. Because the skill writes and versions persona files, accidental mode switches can corrupt profiles or entrench incorrect inferences from sensitive relationship data.

Missing User Warnings

Medium
Confidence
89% confidence
Finding
The script explicitly updates persona and memory artifacts and regenerates `SKILL.md`, but it contains no requirement to warn the user, obtain confirmation, or preserve clear auditability before modifying stored personal-profile data. In a skill built around intimate chat logs, photos, and social media, silent correction-driven rewrites can alter sensitive representations of a real person and make accidental or manipulative changes harder to detect or reverse.

Missing User Warnings

High
Confidence
98% confidence
Finding
The script solicits highly sensitive third-party data such as private chat logs, social media content, and photos with EXIF metadata to construct an AI simulation of another person, but it provides no consent checks, privacy warning, minimization guidance, or limits on collection. This creates a substantial risk of non-consensual surveillance, exposure of intimate communications, and misuse of location and relationship data, especially because the target person may not know their data is being processed at all.

Missing User Warnings

Medium
Confidence
96% confidence
Finding
The skill is explicitly designed to ingest and analyze highly sensitive personal data, including chat logs, social media content, photo metadata, and inferred relationship traits, yet the prompt contains no privacy warning, consent requirement, retention limit, or handling safeguards. In this context, the omission is dangerous because the skill goes beyond simple summarization into profiling, location extraction, behavioral inference, and attachment-style diagnosis of another person, which materially increases privacy, stalking, and misuse risks.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The skill explicitly ingests highly sensitive personal data including chat logs, photos, social media, and inferred psychological attributes, yet the prompt contains no privacy notice, consent boundary, minimization rule, or handling restriction. In this context, the absence of safeguards is dangerous because the skill is designed to build an intimate behavioral model of a third party who may not have consented, increasing risks of privacy invasion, profiling, stalking, and downstream misuse of extracted personal information.

Natural-Language Policy Violations

Medium
Confidence
88% confidence
Finding
The template explicitly instructs the resulting skill to impersonate a real person and 'strictly follow' persona and memory content without any user opt-in or safety boundary. In the context of a skill designed to model a crush from private chats, photos, and social media, this increases the risk of manipulative impersonation, privacy misuse, and emotionally coercive interactions.

Missing User Warnings

High
Confidence
98% confidence
Finding
The prompt explicitly performs psychological profiling on highly sensitive personal data sources without any built-in warning, consent check, or minimization requirement. Because the skill context centers on ingesting a crush's private communications and social media to infer attachment style and vulnerabilities, it normalizes intrusive analysis of a non-consenting person and can facilitate stalking, emotional manipulation, or misuse of intimate data.

Missing User Warnings

Medium
Confidence
90% confidence
Finding
The tool writes derived analysis and selected message content to a filesystem output path without any privacy notice, consent checkpoint, minimization, or protection controls. In this skill context, the inputs are intimate chat logs about a 'crush,' so exporting them to disk increases the risk of unauthorized disclosure, accidental sharing, or later misuse of highly sensitive personal data.

Static analysis

No suspicious patterns detected.