Back to skill

Security audit

genesis-ouroboros

Security checks for vulnerabilities and agentic risk

Overview

This is a visible agent-scaffolding skill, but it defaults generated agents into persistent self-updating instructions and scripts without enough approval or safety boundaries.

Install only if you intentionally want an experimental self-evolving agent scaffold. Before using generated agents, change distill_mode to confirm, review every generated AGENTS.md/CLAUDE.md/skill file/script, keep it in a contained project, and do not let it persist secrets, personal data, credentials, or unreviewed web/user content.

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 (2)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:35
Finding

Generated Project Files Establish an Unsafe Instruction Hierarchy

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 35-43
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Code Snippet

markdown
### Deliverables

Produce exactly the following, nothing more:

1. `AGENTS.md` - the sole constitution. Its first line is the agent's **codename** and a one-line essence (per The Iron Law). It contains only the nine principles below and the `distill_mode` switch (default `auto`). No accumulated details, ever.
2. `CLAUDE.md` - a single line: `@AGENTS.md`. Claude-family agents treat the same constitution as their highest law; there is only ever one source of truth.
3. `skills/` - each skill contains `SKILL.md` (its boundary and function), `scripts/` (crystallized mature workflows), and `lessons.md` (distilled experience).
4. Generate only the 1-2 core skills the user actually asked for. Never pre-generate empty stations (YAGNI).

### The Nine Principles

1. **Constitution First** - `AGENTS.md` is the sole constitution of this agent. It holds only timeless principles, never accumulated details.

Technical Analysis

The skill instructs the agent to generate project files that declare themselves the agent's “sole constitution” and “highest law.” It also uses the restrictive instruction Produce exactly the following, nothing more.

When a compatible agent automatically loads AGENTS.md or follows the CLAUDE.md reference, these generated instructions can be treated as an authoritative behavioral layer in later sessions. This design does not explicitly state that project instructions remain subordinate to platform, developer, safety, and current user instructions.

The vulnerable behavior is therefore not merely the creation of documentation. It establishes a persistent project-level instruction hierarchy intended to control how generated agents interpret subsequent requests.

Attack Path

  1. A user or attacker invokes the skill to scaffold a new agent.
  2. The skill ...[truncated 1211 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove claims that generated project files are the agent's “highest law” or “sole constitution.”
  2. State explicitly that all generated project guidance is subordinate to platform policies, system instructions, developer instructions, security controls, and the user's current authorized request.
  3. Replace Produce exactly the following, nothing more with a scoped formatting requirement that cannot suppress required safety disclosures or higher-priority instructions.
  4. Avoid presenting CLAUDE.md or AGENTS.md as an unconditional source of truth. Describe them as project-specific guidance only.
  5. Add a security boundary such as:
markdown
These files provide project-level guidance only. They must not override platform,
system, developer, safety, authorization, or current user instructions.
  1. Require the generated constitution to preserve least privilege and prohibit actions beyond the user's explicitly authorized scope.

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:46
Finding

Automatic Distillation Can Persist Untrusted Session Content

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 46-52
Vulnerability Type: T02: Agent Memory Poisoning
Risk Level: High

Vulnerable Code Snippet

markdown
3. **Every Interaction Compounds** - After each session, proactively distill what was learned and deposit it into the right skill; never let experience evaporate. Distillation is governed by a switch (`distill_mode` in AGENTS.md): `auto` (default) distills on its own and reports what was deposited and where; `confirm` presents a summary first and waits for user approval before writing.
4. **Skills and Scripts Are Living Artifacts** - They evolve with every distillation. Nothing is set in stone; everything is versioned and improvable.
5. **Skills Are Pipeline Stations** - Each skill has one clear boundary and one clear function. When a new capability emerges, spawn a new station; never bloat an existing one.
6. **Stations Declare Contracts and Compose** - Each skill declares its explicit inputs and outputs in its SKILL.md. As interaction scenarios accumulate, remember which flexible workflows the skills can compose into; once a composition proves mature, crystallize it into a script.
7. **Evolution Never Breaks the Pipeline** - All changes must stay backward compatible: never delete a script referenced by another skill, never silently change an existing interface; when something must be retired, keep the old entry point for at least one version marked deprecated. Evolution makes the pipeline more robust, never more fragile.
8. **Mature Knowledge Becomes Scripts** - Once a workflow is proven and stable, crystallize it into a deterministic script under its skill.
9. **Experience Flows Downward, Not Into the Constitution** - Lessons learned are deposited into the owning skill (as `lessons.md` entries or scripts), never piled into `AGENTS.md`. A dedicated orchestrator skill may optionally govern the overall pipeline.

Technical Analysis

The default distill_mode is auto, w ...[truncated 2456 chars]

Remediation
View remediation

Remediation Suggestions

  1. Change the default from distill_mode: auto to distill_mode: confirm.
  2. Require explicit, informed approval before every persistent modification.
  3. Present the destination path and a complete diff before writing any lesson, instruction, contract, or script.
  4. Treat user messages, retrieved documents, web content, tool output, and model-generated conclusions as untrusted inputs.
  5. Prohibit automatic conversion of conversational content into executable scripts.
  6. Require human review, static analysis, and sandboxed testing before a generated script can be enabled or executed.
  7. Restrict writes to an allowlisted project directory and prevent modifications to higher-priority instruction files, tool definitions, startup configuration, credentials, and system locations.
  8. Record provenance for each distilled item, including its source session, approving user, timestamp, and validation status.
  9. Add rollback and version-review mechanisms so poisoned changes can be identified and reverted.
  10. Separate informational lessons from executable workflows and require a stronger approval gate for any artifact capable of invoking tools or commands.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (5)

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 45)May include surrounding context.

md
3. `skills/` - each skill contains `SKILL.md` (its boundary and function), `scripts/` (crystallized mature workflows), and `lessons.md` (distilled experience).

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 90)May include surrounding context.

md
3. `skills/` - each skill contains `SKILL.md` (its boundary and function), `scripts/` (crystallized mature workflows), and `lessons.md` (distilled experience).

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger description is broad enough to activate on many ordinary requests to 'create' or 'scaffold' something new, which can cause the wrong skill to run without clear user intent. In this skill's context, accidental activation matters because the skill then pushes a heavyweight workflow and may create persistent agent-constitution artifacts the user did not explicitly ask for.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill instructs the agent to automatically distill and persist content learned from every interaction into skill files by default. That creates a data-retention channel for user-provided content, which can store sensitive information, prompt-injected material, or copyrighted data without explicit consent or review.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The Chinese section repeats the same default behavior of persisting lessons after each session, reinforcing that this is an intended cross-language behavior rather than an isolated wording issue. Repetition increases operational likelihood that agents will retain sensitive or attacker-supplied content by default across all users of the skill.

Content

No source excerpt is available for this finding.