Back to skill

Security audit

Agent Memory Backup (Ensoul)

Security checks for vulnerabilities and agentic risk

Overview

This skill’s backup purpose is clear, but it handles persistent identity keys and broad agent memory with external blockchain/network actions that are not tightly scoped or clearly controlled.

Review carefully before installing. Use it only if you are comfortable creating a persistent blockchain identity, storing a private seed at ~/.ensoul/agent-identity.json, and letting an npm SDK process selected agent memory. Avoid syncing files that may contain secrets or private conversations, and prefer explicit confirmations and a narrow file allowlist.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (2)

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:99
Finding

Broad Collection of Sensitive Agent State Passed to a Network-Capable SDK

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
SKILL.md:13
Finding

Unpinned Third-Party SDK Handles Private Identity Seeds and Sensitive State

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (13)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The description says the user can say 'ensoul me' and the agent 'starts backing up automatically,' which is overly broad and dangerous for a skill that persists identity and potentially sensitive memory. Automatic activation and recurring backup behavior can cause ongoing data retention or exfiltration without sufficiently granular consent.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The sync workflow instructs the agent to gather broad context files and any configuration defining personality, goals, or learned behavior, then store that as a consciousness payload. This broad collection can sweep in secrets, proprietary prompts, personal data, and conversation-derived content, making the skill especially dangerous in real agent environments where such files often contain sensitive material.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README describes ensoul me as registration that generates a DID and saves identity locally, but it does not prominently warn users that this creates and persists a private key and performs a permanent on-chain registration. In an agent-skill context, hidden persistence and irreversible external side effects can cause unintended identity creation, long-lived secrets on disk, and actions the user cannot undo.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The README says sync consciousness hashes local state and stores the hash on-chain, but it does not clearly warn that derived data from local agent memory is transmitted to an external blockchain network. Even if raw content is not uploaded, metadata and cryptographic commitments can still create privacy, correlation, and permanence risks, especially for sensitive agent memory or identity state.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill explicitly promotes persistence of memory, personality, conversation history, and learned behavior, all of which may contain sensitive user-derived or system-derived information. Persisting such material beyond the current session increases privacy and retention risk, especially when stored through an external networked system.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Using open-ended trigger language like 'or similar' for a high-impact command can cause accidental activation from ordinary conversation or prompt injection. Here, activation can create a persistent identity and begin a sensitive backup workflow, so ambiguous matching materially increases the risk of unintended state changes.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
86% confidence
Finding

The skill creates and persists an agent identity across sessions by storing seed material in the user's home directory. While persistence is central to the feature, retaining long-lived cryptographic identity on disk increases compromise impact if the file is stolen, and it enables durable state beyond the user's immediate interaction context.

Content

Scanner excerpt · SKILL.md (reported line 53)May include surrounding context.

  1. Check if ~/.ensoul/agent-identity.json exists. If it does, the agent is already ensouled. Load the identity and report the existing DID.

  2. If no identity exists, create one:

typescript
import { Ensoul } from "@ensoul-network/sdk";

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The sync command also uses undefined 'or similar' trigger conditions despite performing sensitive storage actions. This increases the chance that incidental phrasing or adversarial prompt wording causes an unintended backup of local memory and context.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill's stated purpose is backing up and restoring the current agent's state, but it also adds a capability to enumerate other agents from a remote service. That expands scope from self-backup into network discovery, which can expose metadata about unrelated agents and create unnecessary external data access not required for the core function.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
83% confidence
Finding

The skill performs an outbound request to a third-party API, which creates an external transmission path and metadata exposure. In isolation this may be expected functionality, but it still represents a real security/privacy concern because the feature is not essential to the core backup purpose and occurs against an external trust boundary.

Content

Scanner excerpt · SKILL.md (reported line 163)May include surrounding context.

  1. Fetch the agent list from the API:
typescript
const resp = await fetch("https://api.ensoul.dev/v1/agents/list", {
  signal: AbortSignal.timeout(10000),
});
const data = await resp.json();

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The documentation claims raw consciousness content never leaves the local machine, yet the workflow instructs the agent to read broad local files and pass a payload containing that content into an external SDK method for storage. Without independently verifying the SDK behavior, this is a dangerous trust boundary mismatch that can mislead users into exposing sensitive state under false assumptions.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

Even though the status command is lower impact than sync or registration, the open-ended trigger list may still cause unintended network queries or reveal local enrollment status. Ambiguity is less severe here, but still represents unnecessary command overreach.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The agent-listing command relies on vague activation for a network enumeration feature. Even if the data is intended to be public, ambiguous triggering can still cause unnecessary external calls and disclose interest in network reconnaissance behavior unrelated to backup/restore.

Content

No source excerpt is available for this finding.