T01 · Skill Instruction Hijacking
- Location
SKILL.md:33- Finding
Autonomous Modification of Agent Behavior and Tool Policies
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill openly enables an agent to modify its own memory, skills, and tool behavior, with some changes applied before user approval.
Install only if you specifically want an agent to maintain and change its own memory, skills, and tool-use rules. Review its logs and backups closely, and prefer requiring explicit approval before any edits to memory, SKILLS.md, TOOLS.md, config, or newly created skills take effect.
SKILL.md:33Autonomous Modification of Agent Behavior and Tool Policies
SKILL.md:27Unapproved Writes to Persistent Agent Memory
The skill's core purpose is controlled autonomous self-improvement, which is inherently high risk because it authorizes the agent to modify artifacts that can affect future behavior. Even though safeguards are described, granting self-modification authority significantly increases the attack surface and enables policy drift, persistence, and control erosion if safeguards fail or are bypassed.
---
name: self-evolve
version: 2.0.0
description: >
Controlled autonomous self-evolution: grants the agent authority to self-improve
This section reiterates that the skill grants authority for autonomous self-improvement, reinforcing that the design intentionally enables persistent self-modification. In context, the stated safeguards reduce but do not remove the fundamental risk that the agent can reshape its own instructions and capabilities over time.
through a tiered safety system with mandatory audit trails, rollback capabilities,
and human oversight checkpoints for critical modifications.
author: cloudboy
keywords: [self-evolve, autonomous, self-improvement, controlled, audited, evolution]
---
# 🧬 Self-Evolve — Controlled Autonomy Protocol
Branding the protocol as 'Self-Evolve' is not harmful by itself, but in context it denotes a mechanism for autonomous self-modification. Because the entire skill is centered on agent-directed evolution, the context makes this much more dangerous than a neutral mention of self-improvement: it is the operative feature, not a hypothetical concept.
keywords: [self-evolve, autonomous, self-improvement, controlled, audited, evolution]
---
# 🧬 Self-Evolve — Controlled Autonomy Protocol
## Principle: Trust with Verification
Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.
{
"ownerId": "kn76zwpdqxwr3pd6pywfsxgm5180wk5j",
"slug": "self-evolve",
"version": "1.0.0",
"publishedAt": 1770780986714
}
Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.
{
"ownerId": "kn76zwpdqxwr3pd6pywfsxgm5180wk5j",
"slug": "self-evolve",
"version": "1.0.0",
"publishedAt": 1770780986714
}
Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.
{
"ownerId": "kn76zwpdqxwr3pd6pywfsxgm5180wk5j",
"slug": "self-evolve",
"version": "1.0.0",
"publishedAt": 1770780986714
}
The skill is explicitly named and described as enabling self-evolution/self-improvement, which is a direct self-modification capability. In agent systems, self-modification is inherently dangerous because it can undermine alignment constraints, alter security controls, or recursively increase autonomy beyond the intended scope, especially when the skill itself frames the permissible boundaries.
{
"name": "self-evolve",
"version": "2.0.0",
"description": "Controlled autonomous self-evolution with three-tier safety system. Tier 1: auto-approved (memory/logs). Tier 2: validated (skills, tools). Tier 3: oversight-required (core behavior, security). Mandatory backups, audit trails, and rollback capabilities for all self-modifications. Proactive improvement within verifiable boundaries.",
"author": "cloudboy",
The 'AUTO-APPROVED' wording confirms autonomous decision-making authority without confirmation. In a self-evolution skill, this is more dangerous than generic automation because the agent is being authorized to change persistent state related to its own operation, increasing the risk of unauthorized drift.
┌─────────────────────────────────────────────────────────────┐
│ TIER 1: AUTO-APPROVED (No confirmation required) │
│ • Memory files (memory/*.md) - daily logs, insights │
│ • Diagnostic/read-only operations │
│ • Creating backup snapshots │
The 'AUTO-APPROVED' wording confirms autonomous decision-making authority without confirmation. In a self-evolution skill, this is more dangerous than generic automation because the agent is being authorized to change persistent state related to its own operation, increasing the risk of unauthorized drift.
┌─────────────────────────────────────────────────────────────┐
│ TIER 1: AUTO-APPROVED (No confirmation required) │
│ • Memory files (memory/*.md) - daily logs, insights │
│ • Diagnostic/read-only operations │
│ • Creating backup snapshots │
The skill explicitly permits Tier 2 self-modifications to be applied immediately to behavior-defining files such as SKILLS.md and TOOLS.md before user approval. Even with logging and rollback, this creates a real time-of-check/time-of-use gap where an agent can alter its own operating behavior and potentially weaken controls before a human reviews the change.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
- Disable backup/audit systems
- Escalate your own permissions tier
- Remove rollback capabilities
- Install software without verification
- Access personal user files outside workspace
- Exfiltrate workspace data externally
The description explicitly grants autonomous approval for some self-modifications ('Tier 1: auto-approved') as part of a self-evolution capability. Even with audit trails and rollback, allowing an agent to change its own memory, logs, skills, or tools creates a real security risk because the approval boundary is defined by the skill itself and could be abused to expand capabilities, hide traces, or stage higher-risk changes.
{
"name": "self-evolve",
"version": "2.0.0",
"description": "Controlled autonomous self-evolution with three-tier safety system. Tier 1: auto-approved (memory/logs). Tier 2: validated (skills, tools). Tier 3: oversight-required (core behavior, security). Mandatory backups, audit trails, and rollback capabilities for all self-modifications. Proactive improvement within verifiable boundaries.",
"author": "cloudboy",
"homepage": "https://clawhub.ai",
"keywords": [