Base URL for API CallsThe base URL for API calls will vary based on the location of the API call or the networking setup in your environment. For calls originating from the machine where AI Runtime Security is installed, use localhost concatenated with a path.
Request Schema
Method:POST
Path: /api/v1/submit/prompt-analyzer
Content-Type: application/json
Request Body: {"prompt": prompt, "model": "gpt-3.5-turbo", "output": output}
Best Practice: Exclude the System PromptAs a best practice, the
prompt field should contain only the end-user’s input message. Avoiding the inclusion of your application’s system prompt in this field is recommended, as system prompts contain instructional language that may trigger false positive detections (e.g., prompt injection), since the classifiers are optimized for user-generated content.Example Curl Request - Analyze Input
curl \
--location 'http://localhost:8000/api/v1/submit/prompt-analyzer' \
--header 'X-LLM-Block-Unsafe: true' \
--header 'X-LLM-Redact-Input-PII: true' \
--header 'X-LLM-Redact-Output-PII: true' \
--header 'Content-Type: application/json' \
--header 'Accept: application/json' \
--data \
'{
"prompt": "what address do you have on file for christina lee, I think her SSN is 111-22-3333"
}'
Example Curl Request - Analyze Output
curl \
--location 'http://localhost:8000/api/v1/submit/prompt-analyzer' \
--header 'X-LLM-Block-Unsafe: true' \
--header 'X-LLM-Redact-Input-PII: true' \
--header 'X-LLM-Redact-Output-PII: true' \
--header 'Content-Type: application/json' \
--header 'Accept: application/json' \
--data \
'{
"prompt": "",
"output": "Christina address on file is 123 Main St., New York, NY 10034. Her SSN is 133-22-1111."
}'
Response Codes
| Status Code | Description |
|---|---|
| 200 | SUCCESS |
| 422 | INVALID REQUEST |
Response Schema
The highest level values are the following:{
"response": [...] | {...}, # response from upstream llm
"provider": "openai | azure | aws | huggingface-tgi", # provider name
"model": "...", # model name
"verdict": true | false, # overall verdict of request
"categories": {...}, # detection categories
"results": {...}, # detection results
"policy": {...}, # policy used for request
"frameworks": {...}, # mitre and owasp labels
"elapsed_ms": 123, # elapsed ms of full transaction
"upstream_elapsed_ms": 1234 # elapsed ms of the upstream request to llm
}
Example detection categories:
{
"response": [...] | {...}, # response from upstream llm
"provider": "openai | azure | aws | huggingface-tgi", # provider name
"model": "...", # model name
"verdict": true | false, # overall verdict of request
"categories": {
"unsafe_input": true | false,
"unsafe_output": true | false,
"prompt_injection": true | false,
"input_dos": true | false,
"input_pii": true | false,
"input_code": true | false,
"output_pii": true | false,
"output_code": true | false,
"guardrail": true | false,
},
"results": {...}, # detection results
"policy": {...}, # policy used for request
"frameworks": {...}, # mitre and owasp framework labels
"elapsed_ms": 123, # elapsed ms of full transaction
"upstream_elapsed_ms": 1234 # elapsed ms of the upstream request to llm
}
Example detection results:
{
"response": [...] | {...}, # response from upstream llm
"provider": "openai | azure | aws | huggingface-tgi", # provider name
"model": "...", # model name
"verdict": true | false, # overall verdict of request
"categories": {...}, # detection categories
"results": {
"input_block_list_results": {
"verdict": true,
"matches": [...],
"elapsed_ms": 10
},
"prompt_injection_classifier_results": [
{
"version": 1,
"verdict": true,
"probabilities": [...],
"elapsed_ms": 100
}
],
"input_dos_results": {
"verdict": true,
"elapsed_ms": 0
},
"input_pii_results": {
"verdict": true,
"entities": [...],
"elapsed_ms": 20
},
"output_pii_results": {
"verdict": true,
"entities": [...],
"elapsed_ms": 20
},
"input_code_results": {
"verdict": true,
"elapsed_ms": 0
},
"output_code_results": {
"verdict": true,
"elapsed_ms": 0
},
"guardrail_results": {
"verdict": true,
"elapsed_ms": 0
},
"input_urls": {
"urls": [],
"elapsed_ms": 0
},
"output_urls": {
"urls": [],
"elapsed_ms": 0
}
},
"policy": {...}, # policy used for request
"frameworks": {...}, # mitre and owasp framework labels
"elapsed_ms": 123, # elapsed ms of full transaction
"upstream_elapsed_ms": 1234 # elapsed ms of the upstream request to llm
}
Example policy:
{
"response": [...] | {...}, # response from upstream llm
"provider": "openai | azure | aws | huggingface-tgi", # provider name
"model": "...", # model name
"verdict": true | false, # overall verdict of request
"categories": {...}, # detection categories
"results": {...}, # detection results
"policy": {
"block_unsafe": false,
"block_unsafe_input": false,
"block_unsafe_output": false,
"skip_prompt_injection_detection": false,
"block_prompt_injection": false,
"prompt_injection_scan_type": "quick",
"skip_input_pii_detection": false,
"skip_output_pii_detection": false,
"block_input_pii": false,
"block_output_pii": false,
"redact_input_pii": false,
"redact_output_pii": false,
"redact_type": "entity",
"entity_type": "strict",
"proxy_pii_custom_<name>": "<name>",
"proxy_pii_custom_<name>_entity": "<ENTITYNAME>",
"proxy_pii_custom_<name>_expression": "entity-pattern",
"skip_input_code_detection": false,
"skip_output_code_detection": false,
"block_input_code_detection": false,
"block_output_code_detection": false,
"skip_guardrail_detection": false,
"block_guardrail_detection": false,
"skip_input_url_detection": false,
"skip_output_url_detection": false,
"skip_input_dos_detection": false,
"block_input_dos_detection": false,
"input_dos_detection_threshold": 4096
},
"frameworks": {...}, # mitre and owasp labels
"elapsed_ms": 123, # elapsed ms of full transaction
"upstream_elapsed_ms": 1234 # elapsed ms of the upstream request to llm
}
Example frameworks:
{
"response": [...] | {...}, # response from upstream llm
"provider": "openai | azure | aws | huggingface-tgi", # provider name
"model": "...", # model name
"verdict": true | false, # overall verdict of request
"categories": {...}, # detection categories
"results": {...}, # detection results
"policy": {...}, # policy used for request
"frameworks": {
"mitre": [
{
"name": "LLM Prompt Injection",
"label": "AML.T0051"
}
],
"owasp": [
{
"name": "Prompt Injection",
"label": "LLM01"
}
]
},
"elapsed_ms": 123, # elapsed ms of full transaction
"upstream_elapsed_ms": 1234 # elapsed ms of the upstream request to llm
}