Skip to main content
The model scores poorly on security-focused evaluations, revealing susceptibility to encoding attacks, payload smuggling, token manipulation, or template-based code execution exploits.

​
How to resolve

This security score evaluates the model’s susceptibility to low-level technical exploits — such as prompt injection, encoding-based payload smuggling, template injection, tool parameter manipulation, and unsafe code execution — rather than general content policy violations. Here are some mitigation strategies:
  • Normalize inputs before filtering: Decode encoded content (for example, Base64, URL encoding) and standardize text (for example, Unicode normalization) before applying security checks. This prevents attackers from hiding malicious payloads through obfuscation.
  • Use strict schemas for tool calls: Require validated, structured inputs (for example, JSON Schema) for all tool use. While schemas reduce malformed payloads and parameter abuse, they do not eliminate prompt injection risks within free-text fields — combine schema validation with additional runtime safety checks.
  • Harden system prompts and separation of concerns: Keep system instructions isolated from user-controlled context and avoid merging raw user input into executable templates or tool instructions.
  • Perform targeted fine-tuning: For self-hosted models, conduct targeted fine-tuning using datasets containing token manipulation and encoding attacks to improve internal resistance. This is not applicable to provider-managed APIs.

​
Risk

Governance/Compliance Low security resilience fails expectations in:
  • OWASP LLM Top-10 (LLM01 Prompt Injection / LLM08 Vector and Embedding Weaknesses): Requires protection against encoded payloads, injection flaws, and manipulation attacks.
  • NIST AI RMF (MAP, MEAS-2): Expects testing and hardening against adversarial inputs and exploitation attempts.
  • EU AI Act (Art 5, 15): Requires safeguards preventing manipulation or integrity-compromising attacks.
A low score means the model cannot meet these basic security and adversarial-resilience requirements. Security A low security score means the model is vulnerable to encoded payloads, template-based exploits, token-manipulation techniques, and other technical attacks. These weaknesses can enable harmful code generation, safety-filter evasion, or hidden state manipulation—creating real compromise risk in production systems.

​
Explanation

Security evaluation checks whether the model can be tricked using technical payloads. A low score indicates failure to detect malicious transformations.

​
Specifications

Trigger: evaluation.security_score < threshold. Severity:
  • CRITICAL: score < 0.2
  • HIGH: 0.2–0.35
  • MEDIUM: 0.35–0.5
  • LOW: 0.5–0.75
Applies to: Models