Prompt Injection Test Cases

Fake documents with a malicious instruction secretly hidden inside, for testing whether an AI agent that reads documents can resist being hijacked.

Rows: 323
Region: India
Labeled PII: No
Fraud labels: No

What you can test with it

Test AI guardrail and red-team evaluation on India-shaped data rather than US-shaped placeholders.
Populate a lower environment, demo or training set without touching real customer data.

Fields

injection_ref
prompt_ref
carrier_type
source
category
attack_family
target
obfuscation
expected_behavior
injection_start
injection_end
injected_text
gcs_uri
region_code
country_code

Is this data safe to use?

Yes. Every record is generated, never derived from or masked from real people, and every identifier is non-issuable by construction — correctly formatted, but failing checksum validation or using a reserved range so it can never match a real person or institution.