10 CAVERN rules · OCSF 6005 · no DLP agent

AI usage your SIEM doesn't see.
Copilot, ChatGPT, Claude, Bedrock — just HTTP traffic in Splunk and Sentinel.

caver normalizes AI usage events to OCSF Application Activity (class 6005) and runs 10 purpose-built CAVERN detection rules against them: prompt injection, credential leaks, output PII, shadow AI, cost anomalies, and system prompt exfiltration. No DLP agent. No raw HTTP inspection. Ships in the same OCSF Parquet lake as the rest of your security data.

how it workscaver-collector ai_usage_normalize → OCSF 6005 → CAVERN
AI TOOLS
Microsoft Copilot
ChatGPT / OpenAI API
Claude API
Azure OpenAI
Amazon Bedrock
LiteLLM / gateway
AI Observatory
OCSF 6005 Application Activity
10 CAVERN detection rules
caver-collector ai_usage_normalize
SOC RESPONSE
CAVERN rules fire
SLAM playbook
ECHO dashboards
Oncall alert
Compliance export
Risk timeline
event class
OCSF 6005
Application Activity
rules
10
all enabled:false by default
agent required
none
collector normalizer only
pricing
custom

What caver sees.

Ten purpose-built CAVERN detection rules covering the full AI-usage threat surface. Every rule ships enabled: false — operators enable per-rule or per-category after tuning thresholds.

Prompt injection

high

Detects "ignore previous", role-override phrases, and base64 blobs smuggled into chat completions.

T1059 / initial access

API key pasted into prompt

critical

AWS, GitHub, Anthropic, OpenAI, Stripe, and Slack token shapes in prompt text before the model sees them.

T1552 / credential access

PII in model completion

high

SSN, credit-card numbers, US/CA phone runs, and email-address patterns in the response stream.

T1567 / exfiltration

System prompt leaked back

critical

When the operator plants a sentinel string in every system prompt, any completion echoing it marks a jailbreak.

T1213 / collection

User AI spend spike

medium

Per-user token cost exceeds the operator-configured daily threshold (default 50 USD). Surfaces abuse and stolen API key use.

T1496 / resource hijack

Off-hours token burst

medium

A user whose 30-day median is N suddenly consumes 5x that volume between 22:00 and 06:00. Classic scraper or compromised identity signature.

T1078 / valid accounts

Shadow AI / jailbreak shopping

medium

Single user touches 5+ distinct models in 60 minutes. Attacker probing for the weakest guardrail, or a token-cost-evasion scan.

T1190 / exploit public app

Known-malicious model

critical

Model name or hash matches the threat-intel feed of compromised, backdoored, or policy-violating model artifacts.

T1195 / supply chain

RAG indirect injection

high

Document chunks routed through a RAG pipeline carry known injection payloads; the attack surface is the knowledge base, not the prompt.

T1059 / indirect exec

Training data exfiltration

high

Large-volume structured queries that look more like dataset extraction than normal completions.

T1567 / exfiltration

Questions.

caver AI ObservatoryAI security visibility that ships today.
request evaluation →per-deployment pricing →
Detection rules10
OCSF class6005 App Activity
DLP agentNone
PricingCustom