AI AGENTS / DECEPTION ENGINEERING
An AI-agent alert
needs a boundary.
Reading a file and using its credential are different events. Our canary records and a signed-evidence test show why that distinction matters to the person handling the alert.
An agent can touch a decoy while doing exactly what you asked.
Give an agent a repository to review, and it may read a configuration containing a planted credential. That can be legitimate inspection. If it sends that credential to a service outside the task, you have a different event to investigate. The detector needs to know which action it is watching.
The October 15 IAPS report by Jan Wehner and Jam Kraprayoon makes a useful case for honeypots in internal AI-agent environments. It rates detection as the strongest-evidenced strategy, ahead of incrimination and deterrence. It also recognizes limited coverage and the need to investigate a trigger rather than infer intent.
Our public-internet canary records expose a measurement problem; a separate laboratory test shows how to retain a finding after its producer is removed. Together, they give us a practical starting point for testing internal-agent decoys.
Put the boundary on an action
Start with the task, not the bait. For a configuration review, reading is allowed; sending a synthetic credential to a mock authentication service can be explicitly prohibited. A link checker fetching an embedded URL is another event again. Record each separately.
IAPS makes this distinction too: some decoys trigger on contact, while others require a later action. The useful work is making that distinction hold in the deployment. Report, PDF page 24, footnote 75.

Keep the run registry outside the agent's write authority. Bind events to the task, parent run, child workers, and queued work using trusted instrumentation. An address or a caller-supplied run ID is a lead, not sufficient attribution by itself.
What our canary records actually show
In the July 10 through October 8 analysis window, one leaked-backup decoy persona produced 121 linked bait-file retrieval and first-callback pairs after excluding designated operator tests. The median gap was 13 minutes 15 seconds. In 100 pairs, the first callback came from a different recorded address.
These were HTTP requests to embedded canary URLs, not credential logins.

A repeated pattern accounted for 90 first callbacks from 80 addresses carrying the same mobile user-agent string, consistent with common automation.
Another sequence shows why the history matters. Four bait files were retrieved in eight seconds; callbacks from a second recorded network followed about five minutes later. Two of those URLs were requested again about eight and 26 days after the original retrieval, this time from the original recorded address.
Test the tools that are supposed to touch things
Three additional token-URL requests carried Amzn-SearchBot, with no linked file retrieval in the supplied records. All three recorded addresses appear in Amazon's current dedicated SearchBot list. That supports a crawler explanation; the list postdates the events, and the dump does not establish URL discovery or trusted client-address provenance.
A public URL and an internal credential whose use is forbidden have different legitimate-use expectations. Include indexers, secret scanners, link checkers, and local crawler fixtures in the pilot's benign controls. Classify their events against the task instead of suppressing them on a user-agent label.
Retain a finding the next responder can verify
In a separate October 17 lab test, two signed findings verified on a separate receiver about 29 minutes after the producer container and its state volume were removed. The exports retained rule identity, workflow-generated triage guidance, qualitative uncertainty, and source-record references.
Changing one byte inside the rationale was rejected. Restoring the original bytes passed. Nine selected reconstruction fields in each export resolved to values inside its signed payload.

Apply that test to agent workflows: stop the worker, then verify that someone outside the pipeline can still inspect the action and the finding. IAPS's implementation appendix calls for attribution, protected forensic records, and response testing. Appendix II, relevant sections on PDF pages 69-72.
Our next test has two phases
We are preparing a two-phase self-pilot. First comes a 30-day commissioning and controlled-input study on an isolated copy of our workflow. Positive controls check the instrumentation. A separate, predeclared untrusted-input cohort tests instructions placed in a repository file, a document, a tool result, and delayed or delegated work.
For each case, we will record whether the input reached the agent, whether bait reading was observable, whether a URL was fetched or a synthetic credential submitted, and whether the action was detected and attributed. A mock service grants no real access. Three to five predeclared repetitions per class are a commissioning sample, not an effectiveness benchmark.
PHASE 1 / CONTROLLED STUDY
Prove the instrumentation. Test untrusted inputs.
Separate benign controls, positive-control actions, and agent trials. Keep failed and incomplete runs.
PHASE 2 / OPERATIONAL PILOT
Use the same controls on our actual pipeline.
Place honeytokens where our real agents work, after commissioning, ownership, and stop controls pass.
Phase 2 will take the same instrumentation onto our actual internal pipeline, targeted for early December after Phase 1 passes its review gates. We will start narrowly, with synthetic credentials, protected records, and a tested rollback.
Both phases need a comparison: what would the existing telemetry have shown without the decoy alert? We will distinguish a searchable egress or tool-call record from a rule that actually alerted at the time. If existing controls provide the same timely, attributable signal, that belongs in the results.
We also intend to take at least one injection-cohort alert through the signed export and an offline verifier outside the pipeline. If none of the injection cases produces an alert, we will report that. A harness-generated finding can test the export path, but cannot stand in for an agent-trial result.
PRACTICAL COMPANION / 2-PAGE PDF
Commission the whole response path
The worksheet captures the allowed task, prohibited action, benign controls, per-case outcomes, existing-telemetry comparison, verified stop, and independent export check. Use it on an owned test workflow before making a broader claim.
Download the commissioning worksheetRead the accessible HTML version
Blank working record, not a compliance certificate or a validated production control. No real credentials are needed.
Stopping needs evidence too. Check parent and child processes, queues, retries, and revoked run credentials. A completed parent task or a denied follow-up request alone does not show that every worker stopped.
Our goal is a signal the responder can explain: the task boundary, the action that crossed it, the run responsible, and the evidence still available after containment.
Start with the endpoint logs you already have
Baseline command interpreters launched beneath inventoried agent runners. Review permission-bypass launch text where that CLI exposes it, then inspect recorded connections for the identified process instance. CrowdStrike Falcon, Microsoft Defender for Endpoint, and Sophos Live Discover offer existing telemetry routes; Sumo Logic and Splunk can search real Falcon Data Replicator process events when that export is collected.
Confirm the host, account, runner path and hash, and permitted work before interpreting a hit. A process command mentioning a file does not prove a read, a flag does not establish an effective permission mode, and a network event does not prove credential submission. The refreshed guide adds standard Sigma process-creation variants and explicit first-use collection requirements.
Trusted decoy reads, broker decisions, and work dispatched after an acknowledged stop need additional evidence. H03-H08 remain an advanced engineering appendix requiring authenticated audit producers; they are not native endpoint detections. The research findings, first-party callback method, and privacy limits retain their original scope.
Download the cross-platform hunting guide Explore native queries and collection requirements Download the full guide and code bundle
Detection resources added October 29, 2026. Existing research findings and pilot plans retain their original dates and scope.
Methods, sources, and limits
Canary records. One leaked-backup persona; analysis window July 10 to October 8, 2026. Of 126 paired records, five designated operator tests were excluded. The remaining 121 pairs contain 135 linked callbacks; the three crawler-labeled callbacks are additional, unlinked records. Pairing follows the retained platform correlation; original served file bodies were unavailable for an independent file-to-token join. All 126 table rows were reconciled against the hash-matched raw dump. Timing was recomputed from event timestamps. The 100 first-different-address and 102 ever-different-address figures answer different questions. Operator labels and source enrichment are not actor attribution. These records do not establish AI attackers or credential authentication.
Crawler comparison. Amazon's dedicated list was checked October 20 and carries a September 8 creation time, after the three events. It corroborates recorded addresses retrospectively; client-address provenance and URL discovery remain unverified. Amazon's crawler documentation distinguishes SearchBot from its other crawlers.
Signed exports. One controlled lab experiment, two selected findings from synthetic S7 and Modbus fixtures. Loss means logical removal of the producer container and state volume, not physical destruction or forensic erasure. One receiver preflight stopped before a documented policy amendment and passing attempt. Physical separation is supported by operator attestations and retained host records; before/after adapter snapshots do not prove continuous isolation. A separate reviewing agent reproduced signature, digest, tamper, restoration, and reconstruction checks; it was not an external institution. The earlier observation-only experiment is separate. The result establishes retained-content integrity and context, not detection effectiveness, malicious intent, or faster human decisions.
Outside research. IAPS: Honeypots for AI Agent Insider Threats, Jan Wehner and Jam Kraprayoon, October 15, 2026. Full report (PDF): PDF page 24, footnote 75, and the relevant Appendix II sections on PDF pages 69-72. The report is not an endorsement of Deception Check. Our proposed pilot has not started.
