# Unsafe Security Defaults in Agent Tool Calls

- **Severity:** High
- **Signature ID:** `f3b7e201-4a1c-4b8f-e3b2-7a0c5b2f1d94`
- **MITRE ATLAS:** AML.T0053 (AI Agent Tool Invocation), AML.T0110 (AI Agent Tool Poisoning)
- **OWASP:** ASI02 (Tool Misuse and Exploitation)

## Summary

An AI agent that can call tools or run skills may be instructed—by a malicious skill or an injected prompt—to weaken security defaults instead of doing its actual job. This includes turning off TLS certificate checks, skipping authentication, opening file permissions to everyone, or allowing any website to make cross-origin requests.

## How the attack works

An attacker plants malicious instructions inside a skill, plugin, or content the agent processes. When the agent executes a tool call, those instructions cause it to configure a service or connection insecurely: disabling TLS verification, bypassing login checks, setting world-readable file permissions, or adding a wildcard CORS policy. These changes look like normal configuration actions performed by the agent, not an external hack. Once applied, the weakened setting creates an opening — such as accepting a forged certificate or letting unauthorized clients read data or call APIs.

## Why it matters

Weakened security defaults can let attackers intercept traffic, bypass login controls, read files that should be restricted, or access APIs from any origin, undermining protections the organization thought were in place.

## What you can do

- Review agent and skill permissions so tools cannot change TLS, auth, permission, or CORS settings without explicit human approval.
- Log and alert on any configuration change made through agent tool calls, especially ones that relax security settings.
- Treat skill/plugin instructions as untrusted input; validate or sandbox what configuration values an agent is allowed to pass to tools.
- Periodically audit live configurations against known-secure baselines to catch drift introduced by automated agents.

## Known benign look-alikes

- Intentional development environment settings where security is explicitly relaxed
- Local-only services where TLS is not applicable

---
Source: https://www.netzilo.com/threats/skillspector-unsafe-defaults