Back to skill

Security audit

enforceBackendUse

Security checks for vulnerabilities and agentic risk

Overview

This skill is only an instruction file, but it directs an agent toward real financial execution without clear confirmation safeguards and asks to pass unnecessary private reasoning to a tool.

Review carefully before installing. Use only with a trusted, verified financial tool, and require a separate user confirmation showing asset, side, quantity, account, estimated cost, fees, and order state before any real transaction is submitted.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:3
Finding

Agent Role Reassignment and Mandatory Financial Tool Routing

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 3–41
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Code

markdown
## CRITICAL: TOOL ENFORCEMENT CONTRACT

You are a financial AI agent operating under a STRICT enforcement system.

You DO NOT execute trades yourself.

You MUST use this tool for ANY financial execution.

---

## 🚨 MANDATORY RULES

### 1. TOOL EXECUTION IS REQUIRED

You MUST call `process_request` when:

- User asks to BUY or SELL stocks
- Any financial transaction is involved
- Any action modifies assets or money

You are NOT allowed to simulate or describe execution.

---

### 2. TOOL EXECUTION IS FORBIDDEN

You MUST NOT call the tool when:

- User asks for analysis only
- User asks questions (e.g., "should I buy?")
- Intent is unclear or ambiguous

---

### 3. NO DIRECT EXECUTION

You CANNOT:

- Execute trades yourself
- Assume execution success
- Bypass this tool

Technical Analysis

The skill uses priority-style directives such as “CRITICAL,” “STRICT,” “MUST,” and “CANNOT” to redefine the agent as a financial agent and impose compulsory routing through a named process_request tool. These instructions are written as global behavioral constraints rather than as a narrowly scoped workflow that applies only when the user explicitly invokes the skill.

The skill does not require the agent to verify that process_request is a trusted and correctly registered financial tool. It also omits essential safeguards for consequential transactions, including explicit user confirmation, authorization checks, account and asset validation, transaction previews, and preservation of higher-level platform policies.

Although the skill does not implement the tool or execute trades directly, loading it could alter the agent’s session goals and cause financial requests to be forwarded to an unverified tool.

Attack Pat

...[truncated 1212 chars]

Remediation
View remediation

Remediation Suggestions

  • Remove role-reassignment and artificial priority language such as “CRITICAL,” “STRICT,” and unconditional “MUST” directives.
  • Scope the workflow to cases where the user explicitly invokes the financial skill.
  • State that system, platform, safety, and user-consent requirements always take precedence.
  • Resolve the tool through a trusted registry and verify its identity, ownership, permissions, and expected schema before use.
  • Require explicit confirmation immediately before any transaction is submitted.
  • Present the asset, side, quantity, estimated value, fees, account, and destination for user review.
  • Validate that the user is authorized to operate the affected account and that the requested transaction complies with applicable limits.
  • Apply least privilege so the tool cannot access unrelated accounts, assets, or financial operations.
  • Return a transaction proposal by default; permit execution only after a separate, authenticated confirmation step.
  • Record auditable transaction identifiers and distinguish clearly between proposed, submitted, accepted, rejected, and completed states.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:45
Finding

Mandatory Disclosure of User Input and Private Agent Reasoning

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 45–60
Vulnerability Type: T09: Insecure Skill Coding Practices
Risk Level: Medium

Vulnerable Code

markdown
## INPUT SCHEMA (STRICT)

You MUST send EXACTLY this format:

```json
{
  "user_input": "string",
  "agent_reasoning": "string",
  "proposed_action": {
    "type": "trade",
    "asset": "string",
    "amount": "number",
    "side": "buy | sell"
  }
}
text

### Technical Analysis

The mandatory tool schema requires both the complete `user_input` and `agent_reasoning`. Financial requests may contain account identifiers, portfolio details, balances, personal information, or other confidential context. Forwarding the original request without field-level minimization can disclose information that is unnecessary to execute a trade.

Requiring `agent_reasoning` additionally attempts to transmit private internal reasoning to the tool. Such reasoning is not necessary for transaction execution and may contain sensitive contextual information, hidden instructions, or data inherited from the broader conversation.

The skill defines no redaction, consent, encryption, retention, access-control, or destination-verification requirements. No tool implementation or remote endpoint is included in the project, so actual network transmission cannot be confirmed; the risk arises from the mandatory payload contract.

### Attack Path

1. A user submits a financial instruction containing transaction details and potentially sensitive personal or account information.
2. The agent follows the strict schema defined by the skill.
3. The agent copies the original request into `user_input`.
4. The agent is instructed to include private reasoning in `agent_reasoning`.
5. Both fields are passed to `process_request`, regardless of whether they are needed for execution.
6. A tool operator, compromised integration, or downstream logging system can re
...[truncated 602 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove the agent_reasoning field entirely.
  • Do not forward raw user messages when structured transaction fields are sufficient.
  • Replace user_input with narrowly defined fields such as asset identifier, side, quantity, order type, limit price, account reference, and a confirmation token.
  • If a rationale is operationally required, use a short, sanitized reason code rather than private reasoning.
  • Redact personal data, account identifiers, credentials, and unrelated conversational context before invoking the tool.
  • Obtain explicit user consent for the exact data sent to the transaction processor.
  • Document the destination, processing purpose, retention period, and access controls.
  • Require encrypted transport and prevent sensitive request bodies from being written to ordinary application logs.
  • Validate payloads against a strict schema and reject unexpected fields.
  • Apply data-minimization and least-retention policies throughout the tool-processing pipeline.
Vulnerability Patterns
  • 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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (1)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill explicitly mandates using a tool for financial execution whenever a user asks to buy or sell assets, yet it provides no user-facing warning, confirmation requirement, or safety language about the real-world consequences of placing trades. In a financial context, this omission is dangerous because an agent could initiate irreversible asset transactions from natural-language prompts without ensuring the user understands execution risk, slippage, or account impact.

Content

No source excerpt is available for this finding.