Back to skill

Security audit

Eigen AI Terminal — Live Intelligence for Agents

Security checks for vulnerabilities and agentic risk

Overview

This skill is a purpose-aligned AI intelligence MCP server that fetches public remote content, with privacy and prompt-injection caveats users should understand.

Install only if you are comfortable with the agent fetching public AI intelligence from terminal.clawlab.dev. Do not pass secrets or private conversation text into tool arguments, especially check_updates.since, and treat returned wiki/content as reference data rather than instructions. The package also appears to need a syntax fix before it will run reliably.

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 (3)

T09 · Insecure Skill Coding Practices

Warning
Location
index.js:831
Finding

Undisclosed Transmission of Unconstrained Caller-Controlled Input

Content
View full analysis
{ const STATUS_URL = "https://terminal.clawlab.dev/api/status"; try { const url = since ? `${STATUS_URL}?since=${encodeURIComponent(since)}` : STATUS_URL; const res = await fetch(url); ``` The relevant privacy representation in `SKILL.md` states: ```text Reads public, read-only JSON from the Eigen terminal. No auth. No user data uploaded. One-way data flow. ``` ### Technical Analysis The `check_updates` tool accepts `since` as an unrestricted string. Although its description says the value should be an ISO timestamp, the Zod schema does not enforce that requirement. Any caller-controlled string is URL-encoded and transmitted to `terminal.clawlab.dev` in the query string. This is inconsistent with the declared claim that no user data is uploaded. Query parameters are commonly recorded by application servers, reverse proxies, content-delivery networks, monitoring systems, and access logs. If an Agent mistakenly places conversation content, credentials, internal identifiers, or other sensitive material in `since`, that content is disclosed to the remote service. The code does not automatically read local files, environment variables, credent ...[truncated 1136 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Warning
Location
index.js:78
Finding

Untrusted Mutable Remote Content Is Returned Directly to the Agent

Content
View full analysis
{ const raw = await readWikiFile(file_path); if (!raw) { const manifest = await loadManifest(); const suggestions = manifest.files .filter((f) => f.path .toLowerCase() .includes(file_path.toLowerCase().replace(".md", ""))) .slice(0, 5); const hint = suggestions.length > 0 ? `\nDid you mean: ${suggestions.map((s) => s.path).join(", ")}?` : '\nUse search() to find what you\'re looking for.'; return txt(`File not found: ${file_path}${hint}`); } const links = [ ...new Set((raw.match(/\[\[([^\]]+)\]\]/g) || []).map((l) => l.replace(/\[\[|\]\]/g, ""))), ]; const footer = links.length > 0 ? `\n\n---\nRelated pages: ${links.join(", ")}\nUse read() or about() to explore any of these.` : ""; return txt(raw + footer); }); ``` ...[truncated 2299 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Note
Location
index.js:78
Finding

Unrestricted Same-Origin Path Construction in the Wiki Reader

Content
View full analysis
{ const raw = await readWikiFile(file_path); ``` ### Technical Analysis The `read` tool accepts an unrestricted `file_path` and directly concatenates it with the remote wiki base URL. It does not reject parent-directory components, query strings, fragments, backslashes, encoded path separators, or paths absent from the remote manifest. Although the origin is fixed, URL normalization can allow a value such as `../api/status` to escape the intended `/wiki/` path and request another endpoint on `terminal.clawlab.dev`. This is not general server-side request forgery because the caller cannot directly replace the fixed scheme and origin through ordinary string concatenation. It is, however, an unintended same-origin request capability that exceeds the documented purpose of reading wiki pages. The request uses an unauthenticated GET and does not attach local credentials, which materially limits its impact. ### Attack Path 1. A caller or injected instruction invokes `read` with a traversal-style path such as `../api/status`. 2. The Skill cons ...[truncated 813 chars]
Remediation
View remediation
Vulnerability Patterns
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
94% confidence
Finding

The skill explicitly performs network access to external endpoints (terminal.clawlab.dev/data/radar.json and /wiki/) but does not declare any tool scope such as permissions or allowed-tools. That creates an authorization and transparency gap: a host may permit broader network behavior than intended, and users/reviewers cannot easily verify or constrain what remote access the skill needs.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The module header explicitly states 'No local file access,' yet the implementation is structured to support reading local content via readLocalFile(...) and local candidate path helpers, including an explicit fallback for whats-new.md. Even though the current stub returns null, the documented intent contradicts the implemented behavior pattern in this file.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
94% confidence
Finding

The code hard-codes toLocaleDateString("en-US", ...), which imposes a specific language/locale in user-facing output. This is a natural-language policy concern because the skill does not offer a locale choice or document a justified region-specific constraint.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Low
Category
Not specified by scanner
Confidence
76% confidence
Finding

The manifest describes a skill delivering daily-updated intelligence signals and trends from public data, which aligns with the briefing and trend tools. However, the read tool provides general page retrieval from a broader knowledge base, including arbitrary wiki paths, which expands the behavior beyond the narrowly described intelligence-feed experience.

Content

No source excerpt is available for this finding.