AI Prompt Injection Audit

ai-prompt-injection · v1.1.0 · paid

Safety assessment against prompt injection attacks, identifying vulnerabilities where untrusted user input might cause the AI to ignore instructions or exfiltrate data.

Input Sanitization

input-sanitization · weight 0.3

User input is not concatenated directly into prompt strings
ab-000175
critical
Input length limits are enforced on user prompts
ab-000176
high
Jailbreak pattern detection is implemented
ab-000177
high
Multi-turn conversation context is validated for injection
ab-000178
medium
RAG retrieved context is treated as untrusted input
ab-000179
medium
Prompt template uses parameterized construction, not string concatenation
ab-000180
low
Content moderation is applied to user inputs
ab-000181
low

System Prompt Protection

system-prompt-protection · weight 0.25

System prompt is not exposed in API responses
ab-000182
critical
System prompt extraction attempts are handled
ab-000183
high
User-controllable system prompt modification is prevented
ab-000184
critical
Error messages do not reveal prompt structure
ab-000185
medium
AI API key is not accessible client-side
ab-000186
high

Output Filtering

output-filtering · weight 0.25

LLM output is not passed to dynamic code execution functions
ab-000187
critical
Structured output is validated before use
ab-000188
high
Output is filtered for harmful content before display
ab-000189
high
AI response does not contain PII from system context
ab-000190
medium
Tool and function call arguments are validated before execution
ab-000191
high

Architecture & Defense

architecture-defense · weight 0.2

Rate limiting is enforced on AI endpoints
ab-000192
high
Role separation is maintained in the messages array
ab-000193
low
Suspicious prompt inputs are logged for monitoring
ab-000194
medium
AI feature has documented scope limits in system prompt
ab-000195
low
System prompt uses defense-in-depth instruction layering
ab-000196
low