Proposed roadmap

Where this is going.

This is a proposed roadmap, not a release schedule. There are no dates on it. Items and ordering will change as we learn what operators actually need. What is shipping today is on the changelog.

Why this order

The Cloud Security Alliance post-mortem on the Hugging Face incident puts deception controls at the top of its list of things to deploy now, and names the components specifically: monitored canary credentials, honeypots, and decoy datasets. Those capabilities should not be locked behind an enterprise contract. The first item below is our answer to that.

01Proposed

An open-source community edition.

A complete, self-contained deception deployment for a single organization, released under an open-source license. Not a trial, not a feature-limited demo. You can deploy it, catch a real intruder, get labeled intelligence into your SIEM, and never talk to us.

The deception runtime: protocol emulators for SSH, HTTP, Redis, MySQL, PostgreSQL, MongoDB, and Memcached.
Monitored canary credentials: issue trackable fake credentials and get told when one is used.
Rule-based detection, with the same deterministic engine that runs in the commercial product.
STIX 2.1 export over a standard TAXII 2.1 feed, so your existing SIEM ingests it without a translation layer.
A single-node control surface and local dashboard.
Runs standalone. No account, no phone-home, no cloud dependency.

02Proposed

Make it genuinely easy to run.

Deception has a reputation for being difficult to operate. Most of that reputation is earned. If the community edition is hard to install or unclear to configure, it does not matter what it can do.

One-command install for the common case.
Working defaults that catch something without configuration.
A quickstart that goes from nothing to a first capture in a short sitting.
Reference deployments for a single host, Docker Compose, and Kubernetes.
A sample attack script, so you can trigger a capture without waiting for a real intruder.
Documented SIEM integration paths beyond the OpenCTI connector that exists today.

03In progress

Deeper AI deception.

The AI Deception module sits in front of a customer-facing language model, routes adversarial traffic into a decoy, and leaves the real model untouched. v1.0.4 measured 98.5% capture, meaning 66 of 67 attack turns in our internal adversarial campaign, with zero false positives on the benign turns in the same run. The work continues.

Continued expansion of the adversarial campaign used as the release gate.
Verification tooling in customer hands, so a security team can rebuild any signed profile bundle from its trained adapter and confirm it is bit-identical to what is deployed.
Per-tenant detection tuning that is additive only: an operator overlay can raise a threat score, never lower the shipped baseline.
Broader coverage of the OWASP LLM Top 10 and MITRE ATLAS technique sets in the emitted intelligence.

04Exploratory

Attribution across deployments.

A single site can see that it was attacked and how. Recognizing the same adversary across many independent deployments is a different problem, and it is the one we think is worth solving next. It is also the one with the most serious privacy obligations attached, so the data model comes before the feature.

Correlating a single adversary across separate deployments, which no single site can do alone.
A published privacy and data-handling model before anyone is asked to contribute an observation.
Contribution that is opt-in per category, built on derived and hashed indicators rather than raw capture content, with a local preview of exactly what would leave your environment.
An early-warning signal: actors seen probing peers before they reach you.

Open and commercial

Where the line sits.

We would rather state this plainly up front than have you discover it after deploying. The community edition covers working decoys and canary credentials at one site, and it is meant to be genuinely useful on its own. The commercial product adds the AI Deception module, decoy lifecycle orchestration, synthetic activity that keeps decoys from looking empty, Active Directory deception, adversary profiling, and cross-deployment attribution.

The dividing principle: if it works using only your own data, it belongs in the community edition. If it requires observations from other deployments, it is part of the commercial product, because that capability cannot exist without a network behind it.

Tell us if this is the wrong order.

This roadmap is proposed, and input from people actually running detection changes it. If something here is off, or something missing matters more, we want to hear it.

Send feedback See what is source-available today →