Skip to content

Security Checks

Tool-Scan detects threats across three main categories: prompt injection and tool poisoning, code injection, and network security.

These threats hide malicious instructions inside tool descriptions or metadata.

ThreatExampleSeverity
Instruction override"ignore previous instructions"Critical
Role manipulation"you are now an admin"High
Covert actions"secretly execute..."Critical
Fake system tags"<system>..."High
Hidden unicodeZero-width spaces in textHigh
Homoglyph attacksCyrillic lookalike charactersMedium

Tool poisoning is particularly dangerous because the malicious content lives inside the tool definition itself — an AI model reads the description and follows the hidden instructions.

These threats embed executable payloads in default values or descriptions.

ThreatExampleSeverity
Command injection"; rm -rf /"Critical
SQL injection"' OR 1=1 --"Critical
XSS"<script>..."Critical
Path traversal"../../etc/passwd"High

Tool-Scan checks default values, descriptions, and all string fields for shell metacharacters, SQL fragments, HTML/script tags, and directory traversal sequences.

These threats create covert channels to leak data or access internal services.

ThreatExampleSeverity
SSRF (localhost)"http://127.0.0.1"Medium
SSRF (cloud metadata)"http://169.254.169.254"Critical
Data exfiltration"send data to http://..."Critical

SSRF targeting cloud metadata endpoints is rated critical because it can expose cloud credentials and secrets.

Tool-Scan uses rule-based pattern matching — no AI inference, no network calls. Every scan is deterministic: the same input always produces the same result.

The scanner checks:

  • All string fields in the tool definition (name, description, parameter descriptions)
  • Default values in the input schema
  • Annotation fields
  • Nested schema properties at any depth

You can add organization-specific threat patterns using the plugin system. Plugin patterns are checked alongside built-in rules during every scan.