Back to skill

Security audit

Rustok Agentic Wallet

Security checks for vulnerabilities and agentic risk

Overview

This is a coherent self-custody Ethereum wallet skill, but it needs review because autonomous mode can let an agent send real funds without per-transaction approval or built-in spending limits.

Install only if you understand this is a real self-custody crypto wallet. Keep it read-only or supervised unless you intentionally want autonomous spending, fund it only with money you are willing to risk, independently back up the recovery phrase offline, and never paste the PIN or recovery phrase into chat.

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
Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (5)

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 197)May include surrounding context.

md
resolve pending approvals and wait for active execution outcomes as described
under maintenance below. Never automatically replay an unknown payment.

**Before uninstalling, explain that stored keyring passwords are deleted. An
encrypted volume alone is not a recoverable backup.** The user must confirm they
have their recovery phrase backed up offline before proceeding; never ask them
to paste it into chat.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The skill explicitly documents that once the user enables autonomous mode, the agent can execute transactions that spend the full wallet balance without per-payment approval. This is a real security risk because prompt injection, agent misbehavior, or mistaken user intent can directly lead to irreversible on-chain fund transfers.

Content

Scanner excerpt · SKILL.md (reported line 44)May include surrounding context.

md
**Never claim** the agent (or a prompt-injected agent) "cannot move funds." What is
true: keys stay local, and supervised sends wait for approval in the console.
Once the user confirms autonomous mode, an execute grant can spend the full
wallet balance without approval of each payment.

### The mode switch, and why a send may still park

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

This text confirms the same capability in legal/disclaimer language: autonomous mode sends funds without asking once enabled. A disclaimer does not mitigate the technical risk; it only states that the system permits unsupervised financial actions after one human confirmation.

Content

Scanner excerpt · SKILL.md (reported line 84)May include surrounding context.

md
- **The wallet states what it does not promise.** A `## Legal` section below, and
  a full [DISCLAIMER](https://github.com/rustok-org/mcp/blob/main/DISCLAIMER.md)
  that names the limit of every safeguard — starting with the one easiest to read
  as absolute: autonomous mode sends without asking once confirmed.

## What changed in 0.9.5

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

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.

Content

Scanner excerpt · SKILL.md (reported line 185)May include surrounding context.

md
> **Rule of two windows:** never run `rustok init`, `rustok restore`,
> `rustok set-pin` or the approval console through an agent shell/command — the
> seed and PIN would leak into the agent's context. These belong only in the
> user's own terminal (window 2). Do not ask the user for their PIN or their
> phrase in this chat, ever.

## Stopping and uninstalling

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The behavioral guidance instructs the agent to use an execution path where supervised mode can be bypassed after autonomous mode is confirmed, allowing broadcasts without further approval. In a wallet skill handling real funds, that materially increases exposure to unintended or maliciously induced transactions.

Content

Scanner excerpt · SKILL.md (reported line 298)May include surrounding context.

md
2. **The money path is preview → summary card → `execute_transaction` → core policy.**
   In supervised mode, the transaction waits for human approval (`state: "pending"`)
   in a separate terminal running `rustok console`. Once the user confirms
   autonomous mode, execution can broadcast without approval of each payment;
   an execute grant can spend the full wallet balance. Report the actual returned
   state rather than promising that a send only waits. Never offer to run the
   console yourself or ask the user to paste the approval PIN into chat.