Deception Check

FIELD NOTES / CANARY RESPONSE

A canary fired.
What do you check next?

Find where it was placed. Open the log that explains the access. Give the next action an owner.

The alert tells you something touched the canary. Its placement tells you where to investigate.

A canary is a decoy link, credential, or resource designed to alert when it is used. The useful first question is what could have exposed it. A link inside a document points toward document access and sharing records. A planted credential points toward the service where someone might try to use it. The same alert can lead to very different collections.

Our July 10 to October 8, 2026 records show why that context matters: the requests included approved tests, crawler identity claims, and activity still needing investigation. The next log, rather than the alert count alone, changes the response.

Start with the record you actually have

This retained event records an HTTP request to an issued canary URL. We have removed the token value and source address, and assigned the reporting alias C-01.

REDACTED RETAINED RECORDC-01
Observed (UTC)
2026-07-11 14:23:55
Recorded action
HTTP GET to issued token URL

A URL was requested.
Now find where that URL was placed.

The date and action are retained source fields; C-01 is an alias. This record does not show a credential being accepted by another service.

Preserve the original request and its issuance reference together. Keep the observed time, available network and request fields, and the receiving system's response. Keep the actual token in restricted evidence. Use a case alias when sharing the investigation, so a copied link does not create another callback.

Placement decides which log to open

Join the alert to the record that answers the concern. Choose the row that fits the placement; these are collection paths, not a sequence every case must follow.

Pick the next source of evidence
Where the canary belongsOpen nextQuestion to answer
A document or file shareFile access, sharing, and download recordsWho or what could access the containing file or one of its copies?
A fake credentialThe accepting service's authentication and access recordsWas the credential submitted there, and what was the outcome?
Publicly reachable baitAvailable web/proxy records and crawler identity evidenceHow could the URL have been discovered or requested?

Record the source you need, the time range, and who can retrieve it. Preserve short-lived logs promptly. Missing telemetry is a collection gap to assign, rather than a reason to leave the case with no next action.

Explain known activity without hiding the rest

For an approved test, match the expected token, owner, time window, and intended action. Retain the test receipt in the evidence history even when it leaves the operational queue. A familiar source address alone is too broad a reason to suppress an alert.

A crawler name also starts a check. Three requests in our non-test set advertised Amzn-SearchBot. Amazon documents that user agent and publishes its address ranges. Compare the claim with the available address evidence, retain when the check was made, and investigate how the URL became reachable. A current range list is useful context for an older request, not a complete historical identity record.

What the retained requests contained

JULY 10 TO OCTOBER 8, 2026 / HTTP REQUESTS

143requests to issued token URLs
5Scoped tests

Keep the test context and receipt.

138Non-test requests

Three crawler candidates; 135 other requests.

124 distinct tokens in the 138 non-test requests.
Requests and distinct tokens are different units. A review category describes the next investigation task, not a verdict that a request was malicious.

The five documented tests explain their own requests. The three crawler candidates need identity and exposure checks. The other 135 requests remain matters for context and correlation. None of these counts measures unique attackers or successful credential authentication.

Hand off a question someone can answer

A responder should receive more than a token identifier. Give them its placement, the original event, the known test context, and one explicit next action. If the canary was in a restricted document, that action might be to preserve its access and sharing history. If it represented a credential, it might be to check the destination service for an authentication attempt.

RESPONSE WORKFLOW / RECOMMENDED NEXT STEPS

  1. PlacementWhere did the canary belong, and who could encounter it?
  2. Next logWhich access or authentication record will answer the question?
  3. Named ownerWho will collect it, by when, and who receives the result?

An unexplained request to a tightly controlled token can justify opening a case and preserving evidence immediately. Follow your response process for containment when the surrounding evidence warrants it; a crawler review should not delay an urgent collection.

Test the handoff before the next alert

Choose one placement and run an approved test through the whole path. Confirm the alert reaches the intended person, the issuance reference resolves, and the next log is available to that responder. Record what was missing and give the fix an owner. A delivered notification is only one part of that test.

TAKE ONE PLACEMENT THROUGH THE WHOLE RESPONSE

Build a useful canary handoff

The response worksheet records the placement, original event, next evidence source, owner, and next review. Use it for a real investigation or an approved test.

Download the worksheet PDFUse the editable worksheet

For an agent workflow, also define the task boundary and test the separate control that can stop the run. Our AI-agent honeypot field note explains that distinction.

Sources and counting method

Observation window: July 10 through October 8, 2026, UTC. We recounted requests matched to issued tokens in retained Deception Check records: 143 requests, five matched scoped tests, 138 non-test requests, and 124 distinct tokens in that non-test set. Requests without an issued-token match are outside these totals. The three Amzn-SearchBot requests remain crawler-review candidates. C-01 is one actual non-test request, with token and source fields removed for sharing. These request counts differ from the linked retrieval-to-callback pairs examined in our AI-agent article.

The collection paths and handoff steps are our response guidance. Available logs, identity context, and clock accuracy determine what a particular investigation can establish. Technical reference: Amazon's crawler documentation.