Back to skill

Security audit

Mercado Libre

Security checks for vulnerabilities and agentic risk

Overview

This Mercado Libre skill is mostly coherent, but it needs review because it persists marketplace history locally and has conflicting instructions that could allow credentials to be saved in plaintext memory.

Review before installing if you use Mercado Libre for selling, disputes, or API automation. Keep API tokens, passwords, MFA codes, payment details, and session data out of ~/mercado-libre/ memory files; use a secret manager or keychain instead. Periodically review or delete the local memory, watchlist, automation, and dispute logs, and require explicit confirmation before any purchase, listing update, message, inventory change, or other live marketplace action.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
memory-template.md:60
Finding
Credentials May Be Persisted in Unprotected Markdown Memory## Vulnerability Details **File Location**: `memory-template.md:60` **Related Location**: `setup.md:38-46` **Vulnerability Type**: Plaintext sensitive-data storage **Risk Level**: Medium ### Vulnerable Code Snippets `memory-template.md:54-60`: ```markdown ## Key Principles - Keep memory in natural language, not configuration-style key lists. - Store only information that improves future decisions. - Update `last` after each meaningful working session. - Preserve failed decisions and incidents to avoid repeating mistakes. - Never persist credentials unless user explicitly asks. ``` `setup.md:38-46`: ```markdown ## Internal Notes Policy Maintain concise records in `memory.md`: - context and constraints that affect decisions - price and comparison outcomes that should be reused - seller and automation decisions with rationale - dispute incidents and resolution status Use data minimization. Never persist secrets unless user explicitly asks. ``` ### Technical Analysis The instructions permit credentials and other secrets to be persisted when a user explicitly requests it. The designated memory store is `~/mercado-libre/memory.md`, which is described as a Markdown file. No encryption, restrictive filesystem permissions, access controls, redaction process, retention limit, or secret-manager integration is specified. User consent does not make plaintext secret storage secure. This policy also conflicts with `SKILL.md:23`, which requires API tokens or keys to be stored in user-managed secret storage. An agent following the more specific memory instructions could therefore place authentication material directly in the Markdown memory file. ### Attack Path 1. A user, or untrusted content presented as a user instruction, asks the agent to remember a Mercado Libre API token or another secret. 2. The skill interprets explicit consent as authorization to persist the secret. 3. The agent writes the secret into `~/mercado-libre/memory.md`. 4. The plaintext value ...[truncated 798 chars]
Remediation
## Remediation Suggestions 1. Replace the consent-based exception with an unconditional prohibition against storing credentials, API tokens, passwords, MFA codes, payment data, session cookies, and recovery codes in memory files. 2. Store secrets only in an operating-system keychain or dedicated secret manager with encryption at rest and access controls. 3. Persist only non-sensitive secret references, such as a keychain entry name or secret identifier. 4. Apply least-privilege API scopes and separate read-only credentials from credentials capable of marketplace write operations. 5. Add automatic secret detection and redaction before writing any memory or automation log. 6. Avoid recording authorization headers, complete API payloads containing secrets, or sensitive error output. 7. Define credential rotation and revocation procedures for accidental persistence. 8. Apply restrictive permissions to `~/mercado-libre/` as defense in depth, while making clear that permissions do not replace proper secret storage. 9. Align `memory-template.md` and `setup.md` with the secret-storage requirement in `SKILL.md`. Recommended replacement: ```markdown Never persist credentials or secrets in memory files, even when requested. Store authentication material only in user-managed secret storage and retain only a non-sensitive reference to the secret. ```
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (6)

Hidden Instructions

High
Category
Prompt Injection
Content
integration: pending | done | declined

## Context
<!-- User profile, goals, constraints, and preferred decision style -->

## Priority
<!-- Current focus: compare, buy, sell, automate, or dispute -->
Confidence
70% confidence
Finding
Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Hidden Instructions

High
Category
Prompt Injection
Content
<!-- Tracked products, target prices, and urgency notes -->

## Comparisons
<!-- Product A vs B decisions, trade-offs, and final recommendation -->

## Purchase Decisions
<!-- Buy-now vs wait decisions and rationale -->
Confidence
70% confidence
Finding
Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Vague Triggers

Medium
Confidence
93% confidence
Finding
Line L03 says the file should be used when the user wants to 'find, shortlist, or compare products,' which is a very broad invocation description without clear boundaries or exclusions. It does not specify trigger phrases, contexts, or negative examples, increasing the risk of unintended activation for ordinary product-related conversation.

Vague Triggers

Medium
Confidence
93% confidence
Finding
The skill allows an 'always on for Mercado Libre requests' mode without defining clear trigger boundaries, which can cause the agent to activate in broader contexts than the user expects. That creates a risk of unintended commerce-oriented guidance or actions being applied to adjacent conversations, especially where purchasing, selling, or automation decisions may have real financial consequences.

Missing User Warnings

Medium
Confidence
93% confidence
Finding
The instruction to store user context so future sessions start with the same operating model introduces persistent profiling without a clear user-facing notice or consent flow. Even if the stored data is operational rather than secret, retaining decision preferences and marketplace behavior can expose sensitive commercial patterns or exceed user expectations about session boundaries.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The memory policy directs the agent to retain dispute incidents, seller decisions, automation rationale, and price/comparison outcomes, which may contain sensitive financial, behavioral, or conflict-related information. Although it says to minimize data and avoid persisting secrets, it still lacks a clear privacy notice, retention limits, or consent mechanism, making over-collection and unexpected persistence plausible.

Static analysis

No suspicious patterns detected.