To ensure proper AI security monitoring and analysis, you should use Firewall/AI
Security logs for detailed AI-specific threat information when using the Strata Logging Service. These logs provide more detailed information than
standard threat logs.
Strata Logging Service generates the AI security logs when AI
security threats are detected between AI applications and AI models.
Includes detailed threat snippet identification and reporting.
Provide in-depth threat information and reports for different
protection types such as AI model protection, AI application
protection, and AI data protection.
Create an AI security profile and attach a model group with
specific protections to monitor traffic between your AI models,
AI applications, and AI data, and detect threats.
This log type is recommended for platforms with Strata Logging Service for log forwarding, and if you want
detailed AI-specific threat information.
A log is generated for each AI security threat detected between an AI
application and model. The logs are generated for prompt injections, sensitive
data leakage, malicious URLs detected, and AI-generated database queries.
Traffic Details
Traffic Details: Click on a log to view traffic logs showing general
information about Prisma AIRS AI Runtime: Network intercept.
The traffic details logs include:
The AI model name, AI model CSP region name, AI incident type, AI incident
subtype, AI security profile name, and an incident report ID for
troubleshooting.
The logs also include the session ID, source, and destination details.
The AI security threats are categorized into AI incident types and
subtypes:
Incident Type
Incident Subtype
Incident Subtype Details
ai-model-protection
prompt-injection
NA
ai-app-protection
url-security
URL categories detected
ai-data-protection
data-rule
database-security
Data Rule: Name of DLP profile triggered
Database Security: type of database query detected
(Create, Read, Update, or Delete)
model-denied
N/A
latency-block
N/A
AI Security Report
AI Security Report: The AI Security Report tab provides detailed
information on the AI traffic and specific AI threat logs. It includes threat
snippet identification and reporting. The Prisma AIRS
platform supports asynchronous identification of particular content snippets that
trigger security detectors.
Each report includes a unique report ID to help debug
logs.
Model Protection:
The model protection report displays logs for any
detected prompt injection threats. The prompt injection snippet (at most
1000 characters) helps identify the trigger and can include multiple
snippets for complex payloads.
Application Protection:
The application protection report shows URL
categories and the specific URLs in the payload that triggered these
categories
Data Protection
The data protection report lists DLP data patterns
that were triggered and the masked content that caused each pattern to
trigger. These logs are categorized based on low, medium, and
high-severity alerts. The data protection report includes two main
components:
Matched Data Patterns (DLP):
Shows matched Data Loss Prevention (DLP) data patterns
that were triggered.
Display specific snippets of content matching DLP data
patterns. Content is stored and masked based on
Manage → Data Loss
Prevention →
Settings → Sensitive
Data configuration.
Data pattern matches are categorized based on low,
medium, and high Confidence Level.
If multiple SQL queries are detected in the model
response, the report shows up to 10 SQL queries (with
each query limited to 1,000 characters). The queries are
prioritized in the following order: Delete, Update,
Create, and Read actions.
