Humor

• Outcome: Pass
• Latest audit: Jul 26, 2026 · 12:55 PM UTC
• Version: 1.0.3

Overview

This humor skill is a local personalization skill with disclosed persistent memory, but users should know it saves humor reactions and callbacks on disk.

Install only if you are comfortable with a local humor memory under ~/Clawic/data/humor/. Review or delete that folder if you want to clear learned preferences, callbacks, off-limits topics, or saved successful joke wording.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access

Findings (12)

Context-Inappropriate Capability

Medium

• Confidence: 91% confidence
• Finding: The skill explicitly reads and creates persistent local files containing user-specific humor preferences and learned profile data. For a humor-calibration skill, retaining cross-session personal preference data may be more data collection than users reasonably expect, and the silent persistence increases privacy risk if the data is later accessed, repurposed, or retained indefinitely.

Context-Inappropriate Capability

Medium

• Confidence: 93% confidence
• Finding: The skill directs long-term storage of behavioral and preference data including off-limits topics, context flags, attempt history, running jokes, and verbatim successful responses. This creates a richer user profile than needed for a narrowly scoped humor feature, increasing privacy exposure and the chance of sensitive inferences or replay of personal content.

Missing User Warnings

Medium

• Confidence: 95% confidence
• Finding: The skill explicitly instructs the agent to persist user-generated phrases to a local callbacks file without any mention of user consent, disclosure, retention limits, or sensitivity filtering. Even if the content seems harmless, riffs can contain personal details, inside jokes, names, or contextual identifiers that become unexpectedly sensitive once stored and reused later.

Missing User Warnings

Medium

• Confidence: 90% confidence
• Finding: The skill directs the agent to persist behavioral inferences and interaction history in local files under a user directory, but it provides no notice, consent flow, retention policy, or safeguards around what is stored. Even though the data is framed as humor calibration, it still records user behavior and preferences across sessions, creating privacy and profiling risk if stored unexpectedly or accessed by other components.

Missing User Warnings

Medium

• Confidence: 93% confidence
• Finding: The maintenance section explicitly instructs appending interaction events to history files and deleting older entries while preserving derived conclusions in a profile, again without any privacy warning or user-facing control. This makes the persistence more dangerous in context because the system not only stores behavior but also retains inferred preferences after raw data is trimmed, reducing transparency and making deletion incomplete from the user's perspective.

Vague Triggers

Medium

• Confidence: 86% confidence
• Finding: The phrase 'suspends the probing rules: deliver, don't hedge' is an unconstrained instruction to stop asking clarifying questions once a user asks for humor. That is dangerous because ambiguity around audience, target, consent, and medium is common in humor requests, and removing probing can lead the agent to produce harmful, harassing, or professionally risky content without confirming key constraints.

Vague Triggers

Medium

• Confidence: 86% confidence
• Finding: The phrase 'suspends the probing rules: deliver, don't hedge' is an unconstrained instruction to stop asking clarifying questions once a user asks for humor. That is dangerous because ambiguity around audience, target, consent, and medium is common in humor requests, and removing probing can lead the agent to produce harmful, harassing, or professionally risky content without confirming key constraints.

Missing User Warnings

Medium

• Confidence: 95% confidence
• Finding: The instruction to create profile data silently on first use means the user receives no notice that persistent memory is being established. Hidden state creation undermines transparency and consent, and makes even otherwise low-sensitivity preference storage more dangerous because users cannot meaningfully control or audit it.

Missing User Warnings

Medium

• Confidence: 95% confidence
• Finding: The skill explicitly instructs writing humor-reaction evidence to a local profile file without any disclosure, consent gate, or data-minimization guidance. Even if the data seems low sensitivity, it creates persistent behavioral records about a user, which can accumulate into profiling and privacy risk over time.

Missing User Warnings

Medium

• Confidence: 93% confidence
• Finding: The instruction to log failures to a persistent store implies retention of user interaction outcomes without telling the user that their reactions and conversational context may be saved. Negative-reaction logs can still contain behavioral signals and message context, making the absence of disclosure and retention controls a real privacy issue.

Missing User Warnings

High

• Confidence: 98% confidence
• Finding: This section describes building a user-specific behavioral profile based on response patterns, warmth, length, and emoji usage, then storing that pattern in a profile file. That is explicit behavioral profiling without any privacy warning, purpose limitation, or consent mechanism, which is more dangerous than simple logging because it derives persistent inferences about the user.

Ssd 3

Medium

• Confidence: 96% confidence
• Finding: This instruction creates a clear data retention risk because it tells the agent to save conversational artifacts from the user into persistent storage for future use. In a humor skill, users may casually reveal private context while joking, making the persistence less obvious and therefore more likely to capture information the user did not expect to be retained.

VirusTotal

64/64 vendors flagged this skill as clean.

Static analysis

No suspicious patterns detected.