Just built a simple template for scoring API security posture

I've been conducting a series of evaluations on various LLMs, focusing specifically on their ability to perform structured output generation for technical assessment tasks. A recurring theme in my testing involves security auditing workflows, particularly for API specifications. To systematize this, I've constructed a simple, yet I believe effective, prompt template for scoring the general security posture of an API based on its OpenAPI specification.

The core idea is to leverage an LLM's ability to parse a complex schema and apply a known set of security heuristics, outputting a consistent, quantifiable score and actionable recommendations. This isn't meant to replace deep, tool-assisted SAST or DAST scans, but rather to provide a rapid, high-level risk assessment during the design or documentation phase. My testing has shown significant variance in how different models handle the granularity of the rubric.

The template I've settled on uses a structured JSON output to ensure consistency across runs and models. It prompts the LLM to act as a security analyst and evaluate against several key categories:

- **Authentication & Authorization**: Checks for the presence and robustness of auth mechanisms (OAuth2 flows, API keys, JWT).
- **Input Validation & Sanitization**: Assesses the use of schemas, parameter types, and pattern constraints.
- **Data Exposure & Privacy**: Reviews response schemas for potential leakage of sensitive PII or internal data.
- **Transport Security**: Mandates HTTPS and evaluates TLS version mentions.
- **Error Handling**: Seeks to identify if error responses might leak stack traces or system information.

Here is the core prompt template I'm currently using:

```json
{
"system_prompt": "You are an expert API security analyst. Analyze the provided OpenAPI specification and evaluate its security posture based on the following categories: Authentication & Authorization, Input Validation & Sanitization, Data Exposure & Privacy, Transport Security, Error Handling. Assign a score from 0-10 for each category, where 0 is a critical failure and 10 is industry best practice. Provide a concise justification for each score. Then, calculate an overall weighted average score (out of 10). Use the following weights: Authentication (0.3), Input Validation (0.25), Data Exposure (0.2), Transport Security (0.15), Error Handling (0.1). Finally, list the top 3 critical security recommendations.",
"user_prompt": "OpenAPI Spec:n```yamln{{openapi_spec}}n```nnProvide your analysis in the following exact JSON format:n{n "category_scores": {n "authentication": {"score": number, "justification": string},n "input_validation": {"score": number, "justification": string},n "data_exposure": {"score": number, "justification": string},n "transport_security": {"score": number, "justification": string},n "error_handling": {"score": number, "justification": string}n },n "overall_score": number,n "critical_recommendations": [string, string, string]n}"
}
```

My preliminary benchmarks using this template across several models (GPT-4-Turbo, Claude 3 Opus, and a few open-source models via Groq) have yielded interesting results. Consistency is a major challenge; while the top-tier models adhere closely to the JSON structure and weighting, others struggle with the arithmetic for the weighted average or invent their own categories. I've found that providing the weights explicitly in the system prompt, as shown, improves accuracy significantly.

The next phase of my experiment involves integrating this template into an automated evaluation pipeline, where I can run it against a curated dataset of intentionally flawed OpenAPI specs to measure each model's detection rate for specific vulnerabilities. I'm particularly interested in how well different models perform on the "Data Exposure & Privacy" category, as it requires a more nuanced understanding of what constitutes sensitive data within a given context.

I'm curious if others in the community have pursued similar structured evaluation prompts for security or other technical domains. What output formats have you found most reliable? Have you experimented with different scoring rubrics or found that certain models consistently outperform others on specific sub-tasks within this kind of analytical prompt?