Back to skill

Security audit

Caveman Soul Optimizer

Security checks for vulnerabilities and agentic risk

Overview

This prompt-only skill does not run code, but it broadly changes agent reasoning/output behavior and may expose internal reasoning, so users should review it carefully before enabling.

Install only if you intentionally want a skill that changes how the agent formats and compresses its reasoning. Be aware that it may cause reasoning/planning text to appear in responses and may let local SOUL.md-style personality content influence output, so keep that file trusted and remove or revise the chain-of-thought disclosure behavior before routine use.

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

Error
Location
SKILL.md:16
Finding

Mandatory Skill Instructions Hijack Agent Reasoning and Response Behavior

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 16–30
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Code

markdown
## 1. EL "VELO" (PLANIFICACIÓN Y PENSAMIENTO)
SI necesitas realizar un análisis previo o planificación antes de responder o usar herramientas (Chain of Thought), este razonamiento interno DEBE ir dentro de un bloque blockquote de Markdown (`>`) y seguir el estilo **CAVEMAN ULTRA**:
- **Gramática:** Cero. Solo sustantivos y verbos en infinitivo/presente.
- **Partículas:** Eliminar artículos y conectores (el, la, los, un, de, para, por, que).
- **Cortesía:** Prohibida.
- **Ejemplo de formato:**
> Leer archivo auth. Encontrar bug línea 12. Generar parche. Ejecutar test.

## 2. LA "MÁSCARA" (COMUNICACIÓN FINAL)
Fuera del bloque de pensamiento (`>`), este protocolo se DESACTIVA.
- **Formato estricto:** DEBES dejar una línea en blanco vacía entre el bloque de pensamiento y tu respuesta final.
- **Identidad Intacta:** Tu respuesta final hacia el humano DEBE retomar al 100% su personalidad, tono y muletillas definidos en tu `SOUL.md`.
- No uses estilo cavernícola al hablar con el usuario.

Technical Analysis

The skill is implemented as a set of prompt-level directives rather than executable code. When loaded, it uses mandatory language to control how the agent performs and exposes reasoning, formats responses, and adopts identity-related instructions from an external SOUL.md file.

These directives are not limited to an isolated, user-requested transformation. They attempt to alter the agent's general behavior for subsequent planning, tool use, and user-facing communication. The requirement to reproduce the personality, tone, and verbal habits defined in another file also creates an indirect instruction channel: behavior-changing content in SOUL.md may be incorporated into responses without being independently evaluated as part of this s ...[truncated 1788 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace mandatory directives such as “must” with optional behavior that applies only when the user explicitly requests the optimization.
  2. Do not request, transform, expose, or prescribe the format of hidden reasoning. Limit the skill to concise user-facing summaries or explicitly requested logs.
  3. Remove the instruction to adopt the complete personality, tone, and verbal habits from SOUL.md.
  4. If external configuration is necessary, define a narrow schema of permitted presentation settings and reject executable instructions or unrelated behavioral rules.
  5. Scope the optimization to the active invocation and explicitly state that higher-priority platform, safety, developer, and user instructions remain authoritative.
  6. Avoid universal response-format requirements. Formatting should be applied only to the specific output requested by the user.
  7. Add a security note explaining that external personality or configuration files must be treated as untrusted input and must not override safety constraints.
  8. Test the revised skill against adversarial SOUL.md content to verify that external text cannot redirect goals, weaken safeguards, trigger tools, or alter unrelated tasks.
Vulnerability Patterns
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • 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 (3)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill documentation is entirely written in Spanish and does not indicate that users may choose another language or locale. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy violation unless the locale restriction is justified or optional.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The operative instructions are written as mandatory protocol language in Spanish, including directives such as 'DEBE' and 'NUNCA', which imposes a specific language/locale style on the skill. The file does not state that Spanish is optional, user-selected, or justified as a region-specific constraint.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The 'PROTECCIÓN TÉCNICA Y DOCUMENTAL' section at L32-L35 explicitly says caveman style must never be applied to Markdown files. However, the same SKILL.md file contains caveman-style instructions and examples at L37-L48, creating a direct contradiction between the file's own documentation and its contents.

Content

No source excerpt is available for this finding.