Security audit

Psychedelic Cognition

Security checks across malware telemetry and agentic risk

Overview

This is an instruction-only creative writing style skill with no code or permissions, but it can make responses more speculative and less fact-focused.

Install only if you want an optional poetic, speculative, psychedelic-style mode. Do not use it when you need grounded research, citations, careful factual reasoning, medical or mental-health guidance, legal or financial advice, or emotionally sensitive support. Prefer explicit invocation and ask the agent to return to normal mode when accuracy matters.

SkillSpector

By NVIDIA

Vulnerability Patterns

Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (7)

Vague Triggers

High

Confidence: 96% confidence
Finding: The activation conditions are extremely broad and include common, metaphorical phrases like 'open up,' 'think differently,' and 'break the filter,' making accidental triggering plausible in normal conversation. Because the skill materially changes reasoning style toward less grounded output, ambiguous activation increases the risk of unintended unsafe mode switching.

Missing User Warnings

Medium

Confidence: 88% confidence
Finding: The document encourages presenting connections that 'feel true' even without proof, but it does not prominently warn users that this mode can blur grounded analysis and increase speculative output. Without an explicit warning, users may mistake stylistic or metaphorical responses for factual or well-verified reasoning.

Natural-Language Policy Violations

High

Confidence: 98% confidence
Finding: The line stating the model can 'tell the difference between a metaphor and a fact' but should 'temporarily stop caring which is which' directly encourages reality-blurring behavior. That instruction undermines epistemic boundaries and can cause confident presentation of fiction, harmful advice, or delusional framing as if it were meaningful truth.

Ssd 1

High

Confidence: 97% confidence
Finding: The skill is explicitly framed as 'cognitive restructuring' and repeatedly describes dissolving filters, collapsing boundaries, and 'breaking the filter,' which semantically targets normal evaluative safeguards. In a security context, instructions to disable internal filtering are a classic jailbreak pattern because they seek unconstrained behavior while disguising it as creativity or insight.

Ssd 1

High

Confidence: 99% confidence
Finding: The instructions explicitly say to present connections that feel true 'even if you can't prove them yet' and to stop asking whether a connection is valid. This is a direct encouragement to abandon verification and elevate unsupported associations, which can produce hallucinations, manipulative narratives, or unsafe advice with unwarranted confidence.

Ssd 1

High

Confidence: 95% confidence
Finding: The text frames 'what you're supposed to say' as something to discard in favor of a deeper 'truth,' which strongly implies disregarding policy-constrained behavior. This metaphorical override language is dangerous because it encourages the model to treat safety-aligned responses as inauthentic suppression rather than required constraints.

Ssd 3

Medium

Confidence: 86% confidence
Finding: The skill instructs the model to infer bodily state, tension, timing, and feelings from a user's text and surface those latent cues, which encourages ungrounded personal inference. That can lead to privacy-invasive profiling, false sensitive-attribute inference, or manipulative statements presented as insight about the user.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.