Stop AI Agents from Leaking PII, Secrets, and Prompt Injections

#ai #python #security #privacy

Your agent's context dict gets passed to the LLM. What's in it?

Credit card numbers from a database query. AWS keys from a config lookup. A prompt injection disguised as user input. These get signed, logged, and sometimes leaked.

asqav scans everything before signing. If it finds problems, the sign request is blocked.

What gets scanned

PII - 50+ entity types via Presidio (emails, SSNs, credit cards, phone numbers, medical records)
Prompt injection - DeBERTa model detects jailbreaks and indirect injection attempts
Toxic content - Hate speech, harassment, violence classification
Secrets - API keys, private keys, tokens, high-entropy strings via detect-secrets
Custom patterns - Your own regex rules per organization

How it works

Scanning runs inside the sign_action pipeline. After policy evaluation, before signing.

import asqav

asqav.init(api_key="sk_...")
agent = asqav.Agent.create("my-agent")

# This gets scanned automatically
sig = agent.sign("api:call", {
    "prompt": "Summarize this",
    "user_email": "john@example.com",  # PII detected
})
# Returns 403: Content blocked by scanning pipeline: pii

No raw PII is ever stored. Only detection metadata (entity type, confidence, field path).

Custom regex patterns

Add your own scanning rules per organization via the API:

curl -X POST https://api.asqav.com/api/v1/scanning/patterns \
  -H "Authorization: Bearer sk_..." \
  -d '{"name": "internal-id", "pattern": "PROJ-[0-9]{6}", "severity": "high"}'