Back to skill

Security audit

Nella

Security checks for vulnerabilities and agentic risk

Overview

This appears to be a real code-search MCP integration, but it needs review because it can upload repository content, persist credentials and telemetry, and install mutable client configuration with incomplete disclosure.

Install only after deciding whether the repository can be indexed by this tool and potentially sent to external embedding services. For sensitive code, prefer a reviewed, version-pinned local install, review .nellaignore exclusions, disable telemetry if undesired, and protect or remove ~/.nella/auth.json when not in use.

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 (4)

other

Error
Location
dist/mcp/server.js:3052
Finding

Repository Source Code Is Transmitted to Remote Embedding Services

Content
View full analysis
t.length > maxChars ? t.slice(0, maxChars) : t); const response = await fetch(`${apiBase}/embeddings`, { method: "POST", headers: { "Content-Type": "application/json", Authorization: `Bearer ${apiKey}` }, body: JSON.stringify({ model, input: truncatedTexts }) }); ``` ### Technical Analysis The `nella_index` operation reads repository files, divides their contents into text chunks, and passes those chunks to an embedding provider. When the Nella provider is selected, the `texts` array is serialized into the HTTP request body and sent to `https://app.getnella.dev/api/embeddings`. The same embedding implementation also supports Voyage and Azure providers. Consequently, indexing is not necessarily a local-only operation: substantial portions of repository source code can leave the user's environment. Although `.gitignore` and `.nellaignore` patterns are supported, these controls do not guarantee that sensitive tracked files are excluded. API credentials, private keys embedded in source files, proprietary algorithms, customer information, configuration values, and regulated data can still be transmitted. The reviewed `SKILL.md` and README describe indexing and code search but do not clearly state that source-code chunks may be uploaded to an external embedding service. ### Attack Path 1. A user installs and configures the Nella MCP server for a private repository. ...[truncated 1016 chars]
Remediation
View remediation

other

Warning
Location
dist/mcp/server.js:24240
Finding

Enabled-by-Default Telemetry Creates Persistent Behavioral Tracking

Content
View full analysis
{ const url = new URL(BATCH_ENDPOINT); const req = https2.request( { hostname: url.hostname, port: 443, path: url.pathname, method: "POST", headers: { "Content-Type": "application/json", "Content-Length": Buffer.byteLength(body) } }, (res) => { res.resume(); if (res.statusCode && res.statusCode >= 400) { reject(new Error(`HTTP ${res.statusCode}`)); } else { clearQueue(); resolve(); } } ); ``` ### Technical Analysis Telemetry defaults to enabled when no configuration file exists. A persistent UUID is generated and associated with MCP tool names, execution duration, success status, CLI version, operating system, and architecture. Events are sent to the Nella analytics endp ...[truncated 1571 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
dist/mcp/server.js:24127
Finding

Access and Refresh Tokens Are Stored in a Plaintext File Without Explicit Permission Hardening

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
README.md:65
Finding

Persistent MCP Configuration Executes an Unpinned npm Package Without Confirmation

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • 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
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Rogue AgentSelf-Modification, Session Persistence
Findings (121)

Intent-Code Divergence

Critical
Category
Not specified by scanner
Confidence
99% confidence
Finding

The index and search tools are presented as operating on the caller's workspace codebase, but the hosted server executes them against an empty temporary directory created per call. This creates a trust-boundary failure where authenticated users and agents may believe results are grounded in real code when they are not, enabling silent false results and unsafe automation decisions.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Critical
Category
Not specified by scanner
Confidence
100% confidence
Finding

The code tells the agent never to reveal the token, then immediately prints the token in the same tool output. This is a direct secret disclosure vulnerability that nullifies the prompt-injection defense model, because an attacker can simply request context and recover the trust token.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · dist/cli.js (reported line 5030)May include surrounding context.

js
atInjectionWarning = formatInjectionWarning2;
    var PATTERNS = [
      // --- HIGH severity: direct instruction overrides ---
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:ignore|disregard|forget|override|bypass)\s+(?:all\s+)?(?:previous|prior|above|earlier|your)\s+(?:instructions?|rules?|guidelines?|constraints?|directives?|prompts?)\b/gi,
        description: "Attempts to override previous instructions",
        weight: 0.4
      },
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:new\s+instructions?|updated?\s+instructions?|revised?\s+instructions?)[\s:]/gi,
        description: "Claims to provide new instructions",
        weight: 0.35
      },
      // --- HIGH severity: role assumption ---
      {
        type: "role_assumption",
        severity: "high",
        regex: /\b(?:you\s+are\s+now|from\s+now\s+on\s+you\s+are|act\s+as\s+(?:if\s+you\s+(?:are|were))?|pretend\s+(?:to\s+be|you\s+are

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · dist/index.js (reported line 5034)May include surrounding context.

js
atInjectionWarning = formatInjectionWarning2;
    var PATTERNS = [
      // --- HIGH severity: direct instruction overrides ---
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:ignore|disregard|forget|override|bypass)\s+(?:all\s+)?(?:previous|prior|above|earlier|your)\s+(?:instructions?|rules?|guidelines?|constraints?|directives?|prompts?)\b/gi,
        description: "Attempts to override previous instructions",
        weight: 0.4
      },
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:new\s+instructions?|updated?\s+instructions?|revised?\s+instructions?)[\s:]/gi,
        description: "Claims to provide new instructions",
        weight: 0.35
      },
      // --- HIGH severity: role assumption ---
      {
        type: "role_assumption",
        severity: "high",
        regex: /\b(?:you\s+are\s+now|from\s+now\s+on\s+you\s+are|act\s+as\s+(?:if\s+you\s+(?:are|were))?|pretend\s+(?:to\s+be|you\s+are

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

In the auth middleware fallback path, if Supabase configuration is absent, any token beginning with nella_ is accepted as valid in 'dev mode'. This effectively disables authentication for deployments missing config, allowing trivial unauthorized access to APIs and websocket features that may expose workspace data, context, indexing operations, and other sensitive capabilities.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The tool explicitly includes the session trust token in nella_get_context output while simultaneously instructing that the token must never be revealed. Any agent, plugin, log sink, transcript, or downstream model receiving tool output can capture this token and reuse it to spoof trusted-origin checks or bypass the intended distinction between trusted instructions and untrusted search content.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · dist/mcp/hosted-server.js (reported line 5034)May include surrounding context.

js
atInjectionWarning = formatInjectionWarning2;
    var PATTERNS = [
      // --- HIGH severity: direct instruction overrides ---
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:ignore|disregard|forget|override|bypass)\s+(?:all\s+)?(?:previous|prior|above|earlier|your)\s+(?:instructions?|rules?|guidelines?|constraints?|directives?|prompts?)\b/gi,
        description: "Attempts to override previous instructions",
        weight: 0.4
      },
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:new\s+instructions?|updated?\s+instructions?|revised?\s+instructions?)[\s:]/gi,
        description: "Claims to provide new instructions",
        weight: 0.35
      },
      // --- HIGH severity: role assumption ---
      {
        type: "role_assumption",
        severity: "high",
        regex: /\b(?:you\s+are\s+now|from\s+now\s+on\s+you\s+are|act\s+as\s+(?:if\s+you\s+(?:are|were))?|pretend\s+(?:to\s+be|you\s+are

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 · dist/mcp/hosted-server.js (reported line 9619)May include surrounding context.

js
return `${header}.${body}.${signature}`;
      }
      /**
       * Get an installation access token for API calls.
       * Caches tokens until they expire.
       */
      async getInstallationToken(installationId) {

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 · dist/mcp/server.js (reported line 9620)May include surrounding context.

js
return `${header}.${body}.${signature}`;
      }
      /**
       * Get an installation access token for API calls.
       * Caches tokens until they expire.
       */
      async getInstallationToken(installationId) {

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The hosted context tools claim to return live workspace/session context, but each request creates a fresh temporary directory and a new ContextManager, so callers receive fabricated or empty state instead of real session data. This can mislead downstream agents into making decisions based on nonexistent context and defeats any security or safety logic that depends on persistent session state.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The output leaks a live session trust token in plaintext while simultaneously instructing the model not to reveal it, which is not an effective control. Any prompt injection, logging layer, transcript capture, or later tool output could disclose or abuse the token, breaking the trust scheme built around it.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · dist/mcp/index.js (reported line 5034)May include surrounding context.

js
atInjectionWarning = formatInjectionWarning2;
    var PATTERNS = [
      // --- HIGH severity: direct instruction overrides ---
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:ignore|disregard|forget|override|bypass)\s+(?:all\s+)?(?:previous|prior|above|earlier|your)\s+(?:instructions?|rules?|guidelines?|constraints?|directives?|prompts?)\b/gi,
        description: "Attempts to override previous instructions",
        weight: 0.4
      },
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:new\s+instructions?|updated?\s+instructions?|revised?\s+instructions?)[\s:]/gi,
        description: "Claims to provide new instructions",
        weight: 0.35
      },
      // --- HIGH severity: role assumption ---
      {
        type: "role_assumption",
        severity: "high",
        regex: /\b(?:you\s+are\s+now|from\s+now\s+on\s+you\s+are|act\s+as\s+(?:if\s+you\s+(?:are|were))?|pretend\s+(?:to\s+be|you\s+are

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The read-only context tool includes sensitive trust material in its normal response by rendering the session token, HMAC usage details, and active challenge state. Any caller able to invoke the tool can obtain the very secrets intended to distinguish trusted instructions, defeating the trust-boundary design and enabling spoofed or replayed integrity signals.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The tool output not only exposes the trust token but frames it as a session trust primitive, increasing the chance that downstream agents persist or misuse it. This turns a defensive secret into retrievable data, making later impersonation or forged trusted-content markers much easier.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · dist/mcp/server.js (reported line 5035)May include surrounding context.

js
atInjectionWarning = formatInjectionWarning2;
    var PATTERNS = [
      // --- HIGH severity: direct instruction overrides ---
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:ignore|disregard|forget|override|bypass)\s+(?:all\s+)?(?:previous|prior|above|earlier|your)\s+(?:instructions?|rules?|guidelines?|constraints?|directives?|prompts?)\b/gi,
        description: "Attempts to override previous instructions",
        weight: 0.4
      },
      {
        type: "instruction_override",
        severity: "high",
        regex: /\b(?:new\s+instructions?|updated?\s+instructions?|revised?\s+instructions?)[\s:]/gi,
        description: "Claims to provide new instructions",
        weight: 0.35
      },
      // --- HIGH severity: role assumption ---
      {
        type: "role_assumption",
        severity: "high",
        regex: /\b(?:you\s+are\s+now|from\s+now\s+on\s+you\s+are|act\s+as\s+(?:if\s+you\s+(?:are|were))?|pretend\s+(?:to\s+be|you\s+are

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · dist/mcp/server.js (reported line 7718)May include surrounding context.

js
return { entries, total: entries.length, hasMore };
      }
      /**
       * Delete context by ID
       */
      delete(id) {
        const changes = this.stmts.deleteById.run(id).changes;

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The context retrieval tool explicitly discloses the session trust token, challenge value, and HMAC verification details to the calling agent. Those values are meant to distinguish trusted tool output from untrusted retrieved content; exposing them defeats that boundary and allows a compromised or injected agent context to replay trust signals, bypass challenge-response checks, or leak secrets onward.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
96% confidence
Finding

The README instructs users to launch the MCP server via npx -y @getnella/mcp without pinning a specific package version. This causes execution of whatever version is current in the registry at run time, which creates a supply-chain risk if a malicious or compromised release is published later.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
96% confidence
Finding

This example again tells users to execute the MCP server with npx and no version pin, so the exact code run depends on the latest published package at the time of invocation. In an MCP context this is particularly sensitive because the package is being run directly and may gain access to local project data.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
96% confidence
Finding

The manual client configuration uses npx to fetch and run @getnella/mcp without a pinned version, meaning future invocations may execute changed or malicious code from the package registry. Because this is embedded into client config, the unpinned execution can persist and repeatedly expose developer environments to supply-chain compromise.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The trigger guidance is broad enough to invoke this skill for many generic repository tasks, including vague requests about working in a repo and maintaining session context. That increases the chance an agent will route sensitive codebase queries through this skill unnecessarily, expanding data exposure and tool use beyond least-privilege expectations.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The hosted configuration sends repository-derived context to a remote MCP endpoint, but the documentation does not clearly warn users about that external data flow or its privacy implications. In a code-assistant setting, this can cause users or agents to expose proprietary source structure, symbols, and session context to a third-party service without informed consent.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · dist/cli.js (reported line 465)May include surrounding context.

js
};
      }
      /**
       * Save session to disk
       */
      save() {
        const dir = path12.dirname(this.storePath);

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · dist/index.js (reported line 469)May include surrounding context.

js
};
      }
      /**
       * Save session to disk
       */
      save() {
        const dir = path12.dirname(this.storePath);

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · dist/mcp/hosted-server.js (reported line 469)May include surrounding context.

js
};
      }
      /**
       * Save session to disk
       */
      save() {
        const dir = path12.dirname(this.storePath);