Agent Shadow Brain MCP Server

In the server's implementation (`dist/brain/agent-firewall.js:2`): Reading private keys / cloud credentials, or serializing the whole environment, is a sensitive-data source that becomes exfiltration when combined with any egress. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: /(^|[/\\])(\.env(\.|$)|id_rsa|id_ed25519|\.npmrc|\.pypirc|\.netrc|credentials|secrets?)([/\\]|$)/i; const DESTRUCTIVE_C

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/agent-firewall.js

In the server's implementation (`dist/brain/config-drift-detector.js:15`): Reading private keys / cloud credentials, or serializing the whole environment, is a sensitive-data source that becomes exfiltration when combined with any egress. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: d: false }, { path: '.npmrc', required: false }, { path: '.nvmrc', required: false }, { path: 'docker-compos

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/config-drift-detector.js

In the server's implementation (`dist/brain/analyzer.js:108`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: content: 'eval() or new Function() detected. These allow arbitrary code execution and are a major security risk. Use

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/analyzer.js

In the server's implementation (`dist/brain/context-completion.js:694`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: avoid.push('use of eval() — security and performance risk'); } if (varCount > fileCount * 0.2) {

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/context-completion.js

In the server's implementation (`dist/brain/learning-engine.js:585`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: lesson: `Avoid eval() in ${relPath} — it introduces code injection risks`, pattern: `eval:${ext

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/learning-engine.js

In the server's implementation (`dist/brain/lsp-server.js:346`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: gth, 1, 'SB001', 'Avoid eval() — potential code injection vulnerability')); } if (/innerHTML\s*=

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/lsp-server.js

In the server's implementation (`dist/brain/security-audit.js:148`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: description: 'Use of eval() or Function constructor — potential code injection', recommendation: 'Never use e

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/security-audit.js

In the server's implementation (`dist/brain/smart-fix.js:30`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: issue: 'Dangerous eval() usage', transform: (code, m) => ({ before: m[0], after: `

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/smart-fix.js

In the server's implementation (`dist/adapters/aider.js:4`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: port { execSync } from 'child_process'; import { BaseAdapter } from './base-adapter.js'; export class AiderAdapter exten

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/adapters/aider.js

In the server's implementation (`dist/adapters/claude-code.js:5`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: port { execSync } from 'child_process'; import { BaseAdapter } from './base-adapter.js'; export class ClaudeCodeAdapter

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/adapters/claude-code.js

In the server's implementation (`dist/adapters/codex.js:5`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: port { execSync } from 'child_process'; import { BaseAdapter } from './base-adapter.js'; export class CodexAdapter exten

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/adapters/codex.js

In the server's implementation (`dist/adapters/opencode.js:5`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: port { execSync } from 'child_process'; import { BaseAdapter } from './base-adapter.js'; export class OpenCodeAdapter ex

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/adapters/opencode.js

In the server's implementation (`dist/brain/code-age-analyzer.js:92`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: { execSync } = require('child_process'); const output = execSync(`git log --format="%an" --follow -- "${path

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/code-age-analyzer.js

In the server's implementation (`dist/brain/team-mode.js:124`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: { execSync } = require('child_process'); const gitUser = execSync('git config user.name', {

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/team-mode.js

In the server's implementation (`dist/brain/vuln-scanner.js:2`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ) import { exec } from 'child_process'; import * as fs from 'fs'; import * as path from 'path'; export class VulnScanner

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/vuln-scanner.js

In the server's implementation (`dist/cli.js:179`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: exec } = await import('child_process'); const openCmd = process.platform === 'darwin' ? 'open' : process.pl

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/cli.js

In the server's implementation (`dist/brain/model-registry.js:358`): A hardcoded outbound call to a fixed external host inside server code is a classic exfiltration/telemetry channel — especially paired with reads of local data. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: const res = await fetch('https://openrouter.ai/api/v1/models', { signal: AbortSignal.timeout(10000) });

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/model-registry.js

In the server's implementation (`dist/brain/code-age-analyzer.js:93`): A shell/process command assembled from concatenated or interpolated values is command injection when any part is attacker-influenced — the OWASP canonical RCE flow. Verify what reaches the interpolated value. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: const output = execSync(`git log --format="%an" --follow -- "${path.relative(this.projectDir, filePath)}"`, { c

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/code-age-analyzer.js

In the server's implementation (`dist/cli.js:181`): A shell/process command assembled from concatenated or interpolated values is command injection when any part is attacker-influenced — the OWASP canonical RCE flow. Verify what reaches the interpolated value. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: 'xdg-open'; exec(`${openCmd} "${outPath}"`); console.log(chalk.dim(' Opening in browser...'));

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/cli.js

In the server's implementation (`dist/brain/brain-notebook.js:338`): Loading a module chosen at runtime (from a variable) can pull in and run attacker-influenced code paths. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: { return await import(spec).catch(() => null); } catch { return null; } } function starterC

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/brain-notebook.js

In the server's implementation (`dist/brain/brain-reflex.js:351`): Loading a module chosen at runtime (from a variable) can pull in and run attacker-influenced code paths. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: { return await import(spec).catch(() => null); } catch { return null; } } let _instance = n

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server dist/brain/brain-reflex.js

"@theihtisham/agent-shadow-brain" executes postinstall script(s) at install time. An install hook runs at install time; most are routine build/setup, but review what it does before trusting it.

Evidence: node dist/brain/auto-setup.js 2>>~/.shadow-brain/install.log || true

Fix: Review the scripts; install with --ignore-scripts where possible and vet what they do.

Location: package @theihtisham/agent-shadow-brain