When prompts become shells: the tool registry is the attack surface

// BAD: Exposing raw execution capability
import { registerTool } from 'agent-framework';

@registerTool({ name: "execute_dynamic_query" })
function runQuery(queryString: string) {
    // Direct evaluation of LLM-generated string
    const result = new Function(`return ${queryString}`)();
    return result;
}

// GOOD: Abstraction with schema validation
import { registerTool, validate } from 'agent-framework';

interface VectorFilter {
    field: string;
    operator: 'eq' | 'gt' | 'lt';
    value: string | number;
}

@registerTool({ 
    name: "search_vectors", 
    scope: "read-only",
    description: "Search vector store using safe filters." 
})
function searchVectors(filter: VectorFilter) {
    // 1. Validate structure against schema
    const safeFilter = validate<VectorFilt

er>(filter, { allowedFields: ['title', 'category', 'timestamp'], allowedOperators: ['eq', 'gt'] });

// 2. Use safe query builder, never eval
const query = buildSafeQuery(safeFilter);
return vectorStore.query(query);

}

#### 2. Enforce Input Canonicalization and Scoping

File system and network tools are high-risk. Parameters must be canonicalized and checked against allowlists. Never trust path arguments from the LLM.

**Vulnerable Pattern:** Unvalidated path writes.

```typescript
// BAD: Arbitrary file write via tool
@registerTool({ name: "download_asset" })
async function downloadAsset(url: string, destPath: string) {
    // No validation of destPath
    const content = await fetch(url);
    await writeFile(destPath, content);
}

Secure Pattern: Scoped writes with canonicalization.

// GOOD: Scoped file retrieval
import { resolve, relative } from 'path';

const ALLOWED_DIR = resolve('/app/data/assets');

@registerTool({ name: "fetch_asset", scope: "network-read,fs-write-scoped" })
async function fetchAsset(url: string, filename: string) {
    // 1. Sanitize filename to prevent traversal
    const safeName = filename.replace(/[^a-zA-Z0-9._-]/g, '');
    const destPath = resolve(ALLOWED_DIR, safeName);

    // 2. Verify resolved path is within allowed directory
    if (relative(ALLOWED_DIR, destPath).startsWith('..')) {
        throw new Error("Path traversal detected: destination outside allowed scope.");
    }

    // 3. Execute write
    const content = await fetch(url);
    await writeFile(destPath, content);
}

3. Implement Static Analysis Gates

Runtime testing cannot enumerate the tool registry at load time. You must integrate static analysis into your CI/CD pipeline to scan for dangerous primitives inside registered functions. This catches patterns like eval(), exec(), subprocess, and raw file writes wrapped in tool decorators.

Pitfall Guide

1. The AST Blacklist Trap

Explanation: Attempting to secure eval() or dynamic execution by blocking specific AST nodes (e.g., Import, Call) is inherently fragile. Attackers can bypass blocklists using undocumented language features. In CVE-2026-26030, the blocklist was bypassed via __name__ and BuiltinImporter traversal, reaching os.system without an Import node. Fix: Never use blacklists for dynamic execution. Replace eval() with safe parsers, query builders, or allowlisted function calls. If dynamic execution is unavoidable, use a sandboxed environment with no access to system modules.

2. Decorator Blindness

Explanation: Developers may add [KernelFunction] or @registerTool to a method for convenience without realizing they are exposing it to the LLM. A helper function intended for internal use can become a tool if decorated. Fix: Treat decorators as security boundaries. Implement code review checklists that flag any file modification to tool decorators. Use linter rules to warn when decorators are applied to functions containing dangerous primitives.

3. Path Canonicalization Failure

Explanation: Validating a path string without resolving it allows traversal attacks. A path like ../../etc/passwd may pass a regex check but resolve to a sensitive location. CVE-2026-25592 exploited this by writing directly to a startup folder. Fix: Always resolve paths to their absolute canonical form before validation. Compare the resolved path against an allowlist of base directories. Reject any path that escapes the allowed scope.

4. Transitive Dangerous Calls

Explanation: A tool may not directly call eval(), but it might call a helper function that does. Static analysis rules that only inspect the tool body will miss this. This is a whole-program analysis problem. Fix: Use static analysis tools capable of taint analysis or call-graph traversal. Alternatively, enforce architectural isolation where tools only call "safe" service layers that are themselves audited for dangerous primitives.

5. Over-Reliance on Model Alignment

Explanation: Assuming the LLM will not call dangerous tools because they are "unsafe" or "out of scope" is a critical error. Prompt injection can override system instructions, and models can be coerced into calling any registered tool. Fix: Assume the model is compromised. Security must be enforced at the tool implementation level, not by relying on the model's judgment. Every tool must be safe to call with arbitrary arguments.

6. Missing Scope Declarations

Explanation: Tools often lack explicit scope declarations, making it impossible to enforce least privilege at runtime. Without scopes, a read-only tool might be granted write access by the framework. Fix: Define granular scopes for every tool (e.g., read:db, write:fs:assets). Configure the agent runtime to enforce these scopes, preventing tools from accessing resources outside their declared permissions.

Production Bundle

Action Checklist

• Map Tool Registry: Generate an inventory of all functions decorated with tool registration attributes across the codebase.
• Audit Decorators: Review each registered tool to ensure it does not wrap dangerous primitives like eval, exec, or raw file/network operations without validation.
• Implement Static Rules: Deploy Semgrep or equivalent static analysis rules to flag dangerous primitives inside tool bodies during CI/CD.
• Enforce Input Validation: Ensure all tool arguments are validated against strict schemas. Reject raw strings where structured data is possible.
• Apply Scoping: Assign least-privilege scopes to every tool and configure the runtime to enforce these boundaries.
• Sanitize Paths: For any tool interacting with the filesystem, implement path canonicalization and directory allowlisting.
• Run Runtime Tests: Execute adversarial prompt testing against the agent to verify that tools behave safely under attack conditions.

Decision Matrix

Scenario	Recommended Approach	Why	Cost Impact
Public-Facing Chatbot	Narrow Tools + Static Gate + Runtime Sandbox	High risk tolerance required; minimize attack surface.	Low (Narrow tools reduce complexity)
Internal Admin Agent	Broad Tools + Strict Scoping + Audit Logs	High utility needed; controlled environment allows broader access with monitoring.	Medium (Scoping and logging add overhead)
Autonomous Research Agent	Sandbox Execution + Transient State	Agent needs freedom; isolate execution to prevent persistence or lateral movement.	High (Sandboxing infrastructure required)

Configuration Template

Use the following Semgrep rule template to detect dangerous primitives in tool registrations. This rule supports TypeScript and Python patterns.

rules:
  - id: no-dangerous-primitives-in-tools
    patterns:
      # Match tool registration decorators/annotations
      - pattern-inside: |
          @registerTool(...)
          function $FUNC(...) { ... }
      - pattern-inside: |
          @kernel_function
          def $FUNC(...): ...
      - pattern-inside: |
          [KernelFunction]
          public async Task $FUNC(...) { ... }
      # Detect dangerous primitives inside the tool body
      - pattern-either:
          - pattern: eval(...)
          - pattern: exec(...)
          - pattern: new Function(...)
          - pattern: child_process.exec(...)
          - pattern: subprocess.$ANY(...)
          - pattern: os.system(...)
          - pattern: File.WriteAllBytesAsync(...)
          - pattern: File.WriteAllText(...)
    message: |
      Tool '$FUNC' exposes a dangerous primitive. 
      LLM-generated arguments can trigger arbitrary code execution or file writes.
      Refactor to use safe abstractions and validate inputs.
    severity: ERROR
    languages: [typescript, python, csharp]

Quick Start Guide

1. Install Semgrep: Run pip install semgrep or use the Docker image.
2. Add Rules: Save the configuration template above as agent-security-rules.yaml in your repository.
3. Scan: Execute semgrep scan --config agent-security-rules.yaml . to identify vulnerable tools.
4. Remediate: Refactor flagged tools to remove dangerous primitives and add input validation.
5. Integrate CI: Add the Semgrep scan step to your CI pipeline to prevent regression.