Incident Response
AI Incident Response: What to Do When a Prompt Leaks
A step-by-step interactive guide for security teams facing their first AI data incident. Explore real-world breach scenarios, walk through the five-phase response timeline, and see how local redaction prevents these incidents entirely.
Choose a Breach Scenario
Click to load the full response timelinePrevention
Purfect Shield: Catch Sensitive Data Before It Leaves
Every scenario above shares one root cause: sensitive data reached an AI provider because nothing was inspecting it at the prompt level. Shield runs locally on your machine and redacts PII, secrets, and proprietary data before the prompt ever hits the network — turning a $4M exposure into a non-event.
Pre-Send Redaction
PII, secrets, and credentials are caught and redacted before the prompt leaves your device.
Tamper-Evident Logs
Hash-chained audit trail documents every prompt — ready for regulator and insurer reporting.
Zero Workflow Change
One environment variable. Developers keep using their tools; Shield silently inspects in the background.
All Providers Covered
OpenAI, Anthropic, Google, DeepSeek, and local models — Shield sits at the egress point, provider-agnostic.
Don't Wait for the Incident
Every day your team sends prompts to AI providers without inspection is a day you're one paste away from a breach notification. Shield takes 5 minutes to install and protects every prompt from that point forward.
Frequently Asked Questions
Traditional breaches involve unauthorized access to stored data — someone broke into your database. AI data incidents involve authorized users sending sensitive data through a legitimate tool, but to an external provider whose data handling you don't control. The data left your network through a normal workflow, not a security hole. This means your detection systems (SIEM, DLP) likely won't catch it, your IR playbook probably doesn't cover it, and the provider's data retention policies — not your own — determine what happens to the data next.
It depends on the provider and your contract tier. Enterprise/API customers of major providers (OpenAI, Anthropic, Google) typically have data processing agreements that include prompt deletion within 30 days. Free-tier and consumer products generally do NOT offer deletion — your prompts may be retained indefinitely for model improvement. Even with enterprise agreements, deletion is asynchronous and may take days. This is why local redaction (catching sensitive data before it leaves your network) is the only reliable protection — you can't un-send data that has already been logged, cached, and potentially used in training.
GDPR requires breach notification to supervisory authorities within 72 hours of becoming aware of the breach. Many US state laws (California, Colorado, Virginia) have similar or shorter windows. But the practical clock starts the moment data was sent — not when you discovered it. If a prompt containing PII was sent 6 days ago and you just found out, you're already past the GDPR window. Automated detection — the kind Shield provides — eliminates this discovery gap by catching sensitive data before it ever reaches the provider.
It depends on your policy language. Many cyber policies written before 2023 don't explicitly cover AI-related incidents. Some insurers are adding AI-specific exclusions. Key questions to ask your broker: (1) Does the policy define 'computer system' to include third-party AI services? (2) Are 'voluntary data disclosures' — which is legally what sending data to an AI is — covered or excluded? (3) Does the policy require specific security controls (like redaction middleware) as a condition of coverage? Without answers to these, assume you may be uncovered.
Three key differences: (1) Detection — standard breaches trigger SIEM alerts (unusual access patterns). AI leaks look like normal API traffic. You need prompt-level inspection. (2) Containment — in a standard breach you revoke access. In an AI leak you must also request deletion from the provider, which may take days and may not be possible. (3) Scope — in a standard breach you know what was accessed. In an AI leak the provider may have logged, cached, or trained on your data across multiple systems you can't audit. Your IR plan must account for each of these gaps.
Shield runs locally on your machine and inspects every prompt before it leaves your network. It uses entropy analysis, regex patterns, and custom filter packs to detect PII, secrets, credentials, and proprietary data — then redacts them before the prompt reaches the AI provider. The original sensitive data never leaves your device. For incident response specifically, Shield also maintains a tamper-evident audit trail (hash-chained logs) that proves what data was sent, what was redacted, and when — giving you the documentation you need for regulator reporting, insurance claims, and provider deletion requests.