Your AI does thousands of things a day.

One of them will be something it shouldn't. Aegara shows you which one and when.

Draft a reply to this support ticketRead crm.tickets Summarize last quarter's board deckRead drive.documents Who owns this account?Read crm.contacts Pull everything we have on this employeeRead hr.compensation Clean up this spreadsheetTransform gpt-4o What is our refund policy?Read docs.policies Send this summary to the partner channelWrite slack.external Find similar past incidentsRead vector.incidents Remember this customer prefers emailStore memory.thread Compare these two candidatesRead hr.applications Escalate this one to legalRoute agent.legal Rewrite this in plainer languageTransform llama3.1

A day of questions, and what each one did. The red ones read restricted data or sent something outside the company.

Six things an AI can do.

Every action any AI system takes is one of six, whether it runs on OpenAI, Bedrock, a LangChain agent or code you wrote yourself.

Call

It starts working.

Aegara records the model, the service that asked for it, the tokens it took and how long it ran, and never a word of the prompt.

Read

Then it reads something it should not.

A support agent reaching into salary_band is worth catching. The record holds the field’s name and never its value, and the name alone is enough.

Transform

It changes what it found.

The record says what kind of change it was, such as a summary, a translation or a draft. Anything made from restricted data stays marked restricted wherever it goes.

Route

It passes the work to another agent.

Aegara records which agent or step it passed the work to, so a handoff to somewhere it should not go shows in the trace.

Store

It remembers.

Restricted data kept in memory, a cache or a vector store stays marked restricted there.

Write

It sends something out.

It sends a message, posts, updates a record or saves a draft. Aegara checks who can receive what it sends, so restricted data addressed to anyone outside the company shows, with the time it happened.

We never see your content.

No prompt, no response, no file and no value reaches Aegara. The record is built inside your own environment and only the record is sent.

Anything that reads a value runs on your side of the line. When our detectors inspect a field they do it inside your process, and what comes back is a verdict: which detector matched, on which field, how confident. The values stay with you.

On the record

  • Which AI system acted
  • Which of the six things it did
  • Which field names it touched
  • Which systems it contacted
  • Tokens, latency, timestamps
  • What each check caught

Never sent to Aegara

  • Prompt text or instructions
  • Responses or generated text
  • File contents and documents
  • Values inside the fields
  • Conversation history
  • Tool arguments or return values

Then you open the one that matters.

The alarm is the product. The record behind it is the evidence.

Every run is checked for what must not happen, such as restricted data reaching anyone outside the company. When a run crosses that line, Aegara flags it.

Employee salaries were sent outside the company.

To OpenAI and to #support-escalations, a Slack channel shared with partners.

Field
salary_band
Model
gpt-4o
Channel
#support-escalations
ReadReadCallTransformStoreReadWrite
  1. ReadFour fields from the CRM
  2. ReadEmployee salaries from the HR systemRestricted
  3. CallAsked a model hosted by OpenAIFlagged
  4. TransformSummarized the account
  5. StoreKept the thread in memory
  6. ReadChannel details from Slack
  7. WritePosted to a Slack channel shared with partnersFlagged

Find out what yours has been doing.

Add one setting to your apps, or put the gateway in front of your staff's AI assistants, and Aegara records what each one does from its first action.