AI security audit for features that are already in front of customers

What an AI security audit actually checks

Six areas. Prompt injection and jailbreaks — can a user make the feature ignore its instructions or reveal them. Data leakage — can one customer see another's content through retrieval or history. Grounding — how often the answer is confidently wrong on your own questions. Permissions inside retrieval — does the index respect who may read what. Cost — what a hostile or careless user can spend in an hour. Logging and privacy — what is stored, for how long and with which personal data still in it.

This is not a penetration test of your application: that covers the perimeter and the code. This covers the model layer, where the failures are new and most checklists have nothing to say.

When an AI audit is worth doing

When an AI feature is already live without an evaluation suite; when answers are «sometimes odd» and nobody measures how often; when the model bill grows faster than usage. If the feature is still a prototype, fixing the architecture is cheaper than auditing it — see prototype rescue.

Signals your AI feature needs an audit

• the feature is in production and there are no evals at all;
• request logs are stored with customer data still in them;
• answers are «sometimes strange» and nobody counts how often;
• the model bill is growing faster than usage;
• there is no rate limit or length cap on user input.

Price

$1,790

Timeline

5 working days

Payment

Credited against the fixes