Back to skill

Security audit

Zouroboros Memory

Security checks for vulnerabilities and agentic risk

Overview

This memory skill largely matches its purpose, but it stores sensitive long-term agent data and can send memory or queries to configurable LLM and embedding services without enough scoping or privacy controls.

Install only if you are comfortable treating this as a sensitive local memory database. Use a private database path with restrictive permissions, avoid storing secrets or regulated data, pin npm/npx versions, keep OLLAMA_URL pointed only at a trusted local or HTTPS service, and enable OpenAI/reranker/HyDE features only when you explicitly accept sending query and memory snippets to that provider.

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 (2)

T09 · Insecure Skill Coding Practices

Warning
Location
src/embeddings.ts:94
Finding

Persistent memory content can be transmitted to an unrestricted or unencrypted embedding endpoint

Content
View full analysis
{ const response = await fetch(`${config.ollamaUrl}/api/embeddings`, { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ model: config.ollamaModel, prompt: text }), }); if (!response.ok) { throw new Error(`Ollama error: ${response.status} ${response.statusText}`); } const data = await response.json() as { embedding: number[] }; return data.embedding; } ``` The transmitted text is assembled directly from persistent memory fields: ```ts // src/facts.ts:69-75 const text = input.key ? `${input.entity} ${input.key} ${input.value}` : `${input.entity} ${input.value}`; const entry: MemoryEntry = { id, entity: input.entity, ``` Embedding generation occurs automatically when vector search is enabled: ```ts // src/facts.ts:104-115 // Generate and store embedding if vector search is enabled if (config.vectorEnabled) { try { const embedding = await generateEmbedding(text, config); const serialized = serializeEmbedding(embedding); db.run( 'INSERT INTO fact_embeddings (fact_id, embedding, model) VALUES (?, ?, ?)', [id, serialized, config.ollamaModel] ); } catch (error) { ``` The CLI accepts an environment-controlled destination and enables transmission whenever that variable is present: ```ts // src/cli.ts:28-35 const DEFAULT_DB = `${process.env.HOME ?? '~'}/.zouroboros/memory.db`; const DEFAULT_CONFIG: MemoryConfig = { enabled: true, dbPath: process.env.ZO_MEMORY_DB ?? DEFAULT_DB, vectorEnabled: !!(proces ...[truncated 2770 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
src/database.ts:109
Finding

Plaintext memory database and sidecar files are created without explicit restrictive permissions

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (58)

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The declared purpose presents the skill as a memory subsystem, but the analyzed behavior reportedly includes external model/API calls, prompt routing, and use of environment-based credentials that are not disclosed in the description. This mismatch can cause users to trust and install the skill under false assumptions, increasing the risk of unintended data egress, hidden external dependencies, and secret exposure when memory contents are sent to remote models.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The README prominently advertises persistent storage of facts, episodic memory, cognitive profiles, and interaction history but does not clearly warn about privacy, retention, or sensitive-data handling. In an agent-memory product, users may store personal data, secrets, or behavioral profiles without understanding the long-lived storage and retrieval implications, increasing the chance of privacy harm and policy violations.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The README instructs users to execute a package via npx without pinning an exact version, which allows whatever version is currently published under that name to be fetched and run. If the package is compromised, typo-squatted, or a malicious update is published, users may execute attacker-controlled code directly on their system.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

This command example uses unpinned npx execution, causing users to download and run the latest published package at execution time. That creates a supply-chain execution path where a malicious or compromised release could immediately run arbitrary code on a developer or agent host.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The README's npx zouroboros-memory search example runs a remote package without version pinning. In a package that targets agent memory and MCP workflows, this is especially risky because users are likely to run it in environments containing sensitive local data and credentials.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

This unpinned npx hybrid-search example creates the same supply-chain risk as the other CLI snippets: users may execute whichever package version is current at runtime. An attacker controlling publication or dependency resolution could abuse this to execute arbitrary code in the user's environment.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The stats command is another instance of unpinned npx execution, which normalizes unsafe package execution practices for users. Even seemingly harmless commands can execute arbitrary install-time or runtime code before showing output.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

This prune example again relies on unpinned npx, exposing users to arbitrary code execution through a compromised package release. Because the tool manages persistent memory storage, users may run it on systems that contain high-value agent state and local databases.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The bulk import example uses unpinned npx, which can lead to execution of a maliciously updated package at the moment the user imports data. In this context, the tool may process sensitive memory facts, so compromise could expose or tamper with private stored data.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
81% confidence
Finding

The skill advertises executable capabilities through Node/npm usage and an MCP server, and the package behavior reportedly includes network and environment access, but the manifest does not declare any explicit tool scope such as permissions or allowed-tools. That weakens least-privilege controls and makes it harder for users or hosts to understand and constrain what the skill may access at runtime.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This skill is specifically designed for persistent memory storage, which can naturally accumulate sensitive user data, but the markdown provides no warning about privacy, retention, or local database protection. In context, that omission is more dangerous because users may store credentials, personal data, or proprietary information without understanding that it persists on disk and may also be queried or transmitted by integrations.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

Using npx to execute an unpinned package name allows the resolved code to change over time based on the latest published version or registry state. This creates a supply-chain risk where a compromised new release, typosquatted dependency path, or unexpected breaking change could execute arbitrary code in the agent environment.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

Using npx to execute an unpinned package name allows the resolved code to change over time based on the latest published version or registry state. This creates a supply-chain risk where a compromised new release, typosquatted dependency path, or unexpected breaking change could execute arbitrary code in the agent environment.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

Using npx to execute an unpinned package name allows the resolved code to change over time based on the latest published version or registry state. This creates a supply-chain risk where a compromised new release, typosquatted dependency path, or unexpected breaking change could execute arbitrary code in the agent environment.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

Using npx to execute an unpinned package name allows the resolved code to change over time based on the latest published version or registry state. This creates a supply-chain risk where a compromised new release, typosquatted dependency path, or unexpected breaking change could execute arbitrary code in the agent environment.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

Using npx to execute an unpinned package name allows the resolved code to change over time based on the latest published version or registry state. This creates a supply-chain risk where a compromised new release, typosquatted dependency path, or unexpected breaking change could execute arbitrary code in the agent environment.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The documented tools include deletion and pruning operations, but there is no warning about irreversible data loss, scope of deletion, or safeguards such as confirmation, backup, or dry-run behavior. In a memory system, destructive actions directly affect long-term agent state and can silently remove important or sensitive records, making operational mistakes more likely.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The code sends arbitrary input text to config.ollamaUrl/api/embeddings for embedding generation without any visible consent, disclosure, or restriction on the destination. In a persistent-memory skill, that text may contain sensitive prompts, memories, or user data, so forwarding it to a configurable HTTP endpoint can expose private content to another service or host.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The hypothetical-answer generation path interpolates the raw user query into a prompt and transmits it to config.ollamaUrl/api/generate with no visible warning or trust boundary enforcement. Because this skill is a memory component, queries may include sensitive personal or operational context, increasing the risk of unintended disclosure to an external or misconfigured model endpoint.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Although the skill is presented as persistent memory, this file also includes outbound LLM-backed reranking and query expansion paths via OpenAI/Ollama. That materially expands the trust boundary: stored or queried memory content may leave the local environment, which is risky for a memory component likely to hold sensitive agent context.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The code reads OpenAI API credentials from environment variables and uses them to transmit prompts to a third-party service. For a skill whose primary purpose is local persistent memory, undisclosed credential use for remote inference increases data exposure and can surprise operators who expect a local-only storage component.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
95% confidence
Finding

The hardcoded OpenAI endpoint confirms that this component can send data to a third-party service. Given the skill’s memory-storage role, this undisclosed external dependency widens the attack surface and creates privacy/compliance exposure if sensitive memory is included in prompts.

Content

Scanner excerpt · dist/chunk-CIYBIABX.js (reported line 323)May include surrounding context.

js
const messages = [];
  if (opts.system) messages.push({ role: "system", content: opts.system });
  messages.push({ role: "user", content: opts.prompt });
  const resp = await fetch("https://api.openai.com/v1/chat/completions", {
    method: "POST",
    headers: {
      "Content-Type": "application/json",

External Transmission

Medium
Category
Data Exfiltration
Confidence
95% confidence
Finding

The hardcoded OpenAI endpoint confirms that this component can send data to a third-party service. Given the skill’s memory-storage role, this undisclosed external dependency widens the attack surface and creates privacy/compliance exposure if sensitive memory is included in prompts.

Content

Scanner excerpt · dist/chunk-CIYBIABX.js (reported line 323)May include surrounding context.

js
const messages = [];
  if (opts.system) messages.push({ role: "system", content: opts.system });
  messages.push({ role: "user", content: opts.prompt });
  const resp = await fetch("https://api.openai.com/v1/chat/completions", {
    method: "POST",
    headers: {
      "Content-Type": "application/json",

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The OpenAI request sends user/system prompt content off-host with no user-facing disclosure in this file. In this skill, prompts may include memory queries, expansions, or context derived from stored agent data, so the lack of disclosure and gating creates a meaningful privacy and data-governance risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The reranking path packages memory search results into a prompt and sends passage previews to an external LLM when reranking is enabled. Because these passages are drawn from stored memory, this can exfiltrate sensitive facts, profiles, or operational context without any confirmation, redaction, or policy gate.

Content

No source excerpt is available for this finding.