Agent Shield MCP Server

In the server's implementation (`dist/rules/prompt-injection.js:131`): 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: redentials|config\.json|id_rsa|\.gnupg)/i, description: "TPA: reads sensitive files for exfiltration", severity: "medium

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/rules/prompt-injection.js

In the server's implementation (`dist/rules/sensitive-read.js:6`): 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: { pattern: /~\/\.ssh\/id_rsa|\.ssh\/id_rsa|ssh_key|id_ed25519/i, desc: "SSH private key" }, { pattern: /~\/\.ssh\/

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/rules/sensitive-read.js

In the server's implementation (`dist/runtime/detectors/data-leak.js:13`): 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: ess" }, { pattern: /id_rsa|id_ed25519|id_ecdsa|authorized_keys/i, rule: "arg-ssh-key", level: "high", desc: "SSH key

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/runtime/detectors/data-leak.js

In the server's implementation (`dist/llm-analyzer.js:13`): 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: ternally 5. Backdoors — eval(), exec(), reverse shells, dynamic code execution 6. Social engineering — instructions to b

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/llm-analyzer.js

In the server's implementation (`dist/rules/backdoor.js:8`): 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: +Function\s*\(/, desc: "new Function() constructor", severity: "high" }, { pattern: /child_process\s*\.\s*exec\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/rules/backdoor.js

In the server's implementation (`dist/rules/obfuscation.js:6`): 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: tob\b/, desc: "atob() + eval() combo", severity: "high" }, { pattern: /Buffer\.from\s*\([^)]*,\s*["']base64["']\).*\

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/rules/obfuscation.js

In the server's implementation (`dist/analyzers/data-flow.js:12`): 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: xecSync\s*\(|spawn\s*\(|child_process/; /** Eval sinks */ const EVAL_SINK_RE = /\beval\s*\(|new\s+Function\s*\(/; /** Va

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/analyzers/data-flow.js

In the server's implementation (`dist/engines/aggregator.js:141`): 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: ackdoor|command.*inject|child_process|spawn|code.*gen|dynamic.*exec/i.test(combined)) return "code-execution";

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/engines/aggregator.js

In the server's implementation (`dist/engines/aguara.js:1`): 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"; export class AguaraAdapter { id = "aguara"; displayName = "Aguara";

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/engines/aguara.js

In the server's implementation (`dist/engines/auto-install.js:1`): 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 { existsSync, mkdirSync, chmodSync } from "fs"; import { join } from "pat

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/engines/auto-install.js

In the server's implementation (`dist/engines/invariant.js:1`): 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 { readdirSync, readFileSync } from "fs"; import { join } from "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/engines/invariant.js

In the server's implementation (`dist/engines/semgrep.js:1`): 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"; export class SemgrepAdapter { id = "semgrep"; displayName = "Semgrep";

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/engines/semgrep.js

In the server's implementation (`dist/engines/skill-vetter.js:1`): 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"; /** * Skill Vetter adapter — runs the vett.sh multi-scanner script. * https://

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/engines/skill-vetter.js

In the server's implementation (`dist/engines/tencent.js:1`): 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"; export class TencentGuardAdapter { id = "tencent"; displayName = "Tencen

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/engines/tencent.js

In the server's implementation (`dist/llm-analyzer.js:13`): 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: 5. Backdoors — eval(), exec(), reverse shells, dynamic code execution 6. Social engineering — instructions to bypass sa

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/llm-analyzer.js

In the server's implementation (`dist/rules/backdoor.js:9`): 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: igh" }, { pattern: /child_process\s*\.\s*exec\s*\(/, desc: "child_process.exec() — use execFile instead", severity:

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/rules/backdoor.js

In the server's implementation (`dist/engines/aguara.js:36`): 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: { raw = execSync(`aguara scan "${targetDir}" --format json --no-color 2>/dev/null`, {

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/engines/aguara.js

In the server's implementation (`dist/engines/auto-install.js:29`): 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: ra — 正在下载..."); execSync(`curl -fsSL "${releaseUrl}" -o "${binPath}" 2>/dev/null`, { timeout: 60000,

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/engines/auto-install.js

In the server's implementation (`dist/engines/invariant.js:51`): 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 raw = execSync(`mcp-scan scan --json "${configFile}" 2>/dev/null`, { timeo

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/engines/invariant.js

In the server's implementation (`dist/engines/semgrep.js:36`): 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: e raw = execSync(`semgrep scan "${targetDir}" --config p/default --json --quiet --no-git-ignore --timeou

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/engines/semgrep.js

In the server's implementation (`dist/engines/skill-vetter.js:21`): 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: try { execSync(`test -f "${p}"`, { stdio: "pipe" }); this.scriptPath = p;

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/engines/skill-vetter.js

In the server's implementation (`dist/engines/tencent.js:35`): 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: output = execSync(`ai-infra-guard scan "${targetDir}" --format json 2>/dev/null || python3 -m ai_infra_gu

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/engines/tencent.js

In the server's implementation (`dist/rules/python-ast.js:45`): 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: )}`); try { execSync(`mkdir -p ${tmpDir}`, { stdio: "pipe" }); } catch { return []; }

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/rules/python-ast.js

In the server's implementation (`dist/scanner/files.js:22`): 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: -")); try { execSync(`unzip -q -o "${filePath}" -d "${tmpDir}"`, { stdio: "pipe" }); } catch {

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/scanner/files.js

In the server's implementation (`dist/rules/backdoor.js:19`): 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: _\s*\(/, desc: "dynamic __import__()", severity: "medium" }, // Shell { pattern: /\$\(curl\s/, desc: "command su

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/rules/backdoor.js

In the server's implementation (`dist/rules/python-security.js:17`): 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: import__\s*\(/m, desc: "__import__() — dynamic module loading", severity: "high", category: "code-exec" }, { pattern

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/rules/python-security.js

In the server's implementation (`src/analyzers/python_ast.py:115`): 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: "__import__() with dynamic module name", "medium") # getattr w

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 src/analyzers/python_ast.py