Product Updates

Anomaly Detection: behavioural threat detection built for lean teams

Written by Austin Mitchell | Nov 18, 2026, 10:32:08 AM

Your alert rules already catch what you know to look for. The harder question is how you find what nobody wrote a rule for. A modern SIEM ingests an enormous volume of logs, and attack activity is a small number of needles in that haystack. Slow, quiet activity, or a legitimate account behaving slightly out of character, rarely trips a signature.

 

Anomaly Detection, a new add-on for Guardsix SIEM, shrinks the rest of the haystack to the part where the unknown needle probably is. It runs entirely on-premises, and every anomaly shows its reasoning, so a lean team in a regulated organisation can act on it and explain it to an auditor.

What's new:

  • Anomaly Detection add-on: behavioural baselines surface unusual user and entity activity, including what no rule covers, entirely on-premises with no cloud service and no new server.
  • Explainable by design: every anomaly states what was observed, the baseline, why it scored, and the score.
  • Detection you own: vendor rule templates to start fast, your own rules as readable YAML, and configuration APIs to manage them as code.
  • One detection system: 622 rules, curated out-of-the-box analytics mapped to MITRE ATT&CK, and a searchable use case library, now joined by anomalies in the same alert and incident stream.
  • An updated Log Collection Agent: a single source of events for the detection stack and Anomaly Detection, with file integrity and Windows Registry monitoring for visibility into important change events. 

 

The result is less for your team to look at and nothing hidden from it, with Anomaly Detection staying on your site from first log to final score.

 

Anomaly Detection is the direct result of our partner co-creation programme. RENK, a German industrial and defence manufacturer, worked with Guardsix as a design partner through onboarding, helping shape how the engine learns a baseline and what "explainable" needed to mean in practice for a team running it day to day.

Anomaly Detection: find what no rule was written for

Anomaly Detection learns what normal looks like for each user and entity, then flags what deviates from it. Your analysts and hunters get a prioritised starting point, scored and held in a searchable repository rather than an alert queue.

 

Underneath, it is a rule-based, baseline-driven engine inside Guardsix SIEM, so there is no cloud service, no external model call and no new server. Baselines are built per entity, per peer group and across the organisation, from stored logs or from live ingest. An entity with thin history falls back to its peer group, and behaviour that becomes normal stops being flagged.

 

Peer groups, imported from Active Directory, LDAP or a CSV file, give each baseline context and keep noise down.

 

Learn how Anomaly Detection helps lean security teams catch suspicious behaviours and explain their findings.

Four detectors, each answering one question

Four detectors ship with the new add-on:

 

  • Time window deviation: is this an unusual time for this entity?
  • Entity attribute rarity: has this entity done this before?
  • Entity frequency drift: is this happening more or less often than usual?
  • Statistical measure: is this number unusually large or small?

 

Together they surface signals such as an administrator account active at 03:00 when it never is, a user signing in from a country they have never used, a burst of failed authentications over 15 minutes, or a slow rise in file transfers over weeks. Detectors are modular, so more can ship without waiting for a full SIEM release.

Explainable, and yours to tune

A score without a reason is hard to act on and harder to explain to an auditor. Every anomaly states what was observed, what the baseline was, why it scored, and the score itself, so an analyst decides on real information.

 

Rules are readable YAML, mapped to MITRE ATT&CK techniques. Each rule moves from Draft through Learning and Testing to Production, and only Production rules escalate to incidents. You can simulate a rule against historical logs, then tune thresholds, scope and suppression before it goes live.

 

Nobody starts from a blank page. Curated vendor rule templates install as packs, and you can use them as shipped or clone them into rules of your own that keep a link to upstream updates. Configuration APIs manage rules and templates programmatically, so detection engineers can keep anomaly rules in Git beside the rest of their detection content.

 

Some of what Anomaly Detection surfaces will be benign, and that is expected. The value is that the set you review is dramatically smaller and far richer than raw logs.

A threat detection stack built as one system

Detection-as-code means every rule in Guardsix SIEM is a structured, versioned file rather than a packaged binary: a name, a description, a persistent rule ID, and the query itself, all readable and trackable on their own. The Use Case Navigator is how a team finds its way through that library, full-text search plus a MITRE ATT&CK coverage heatmap, so "do we have anything for this?" is a lookup, not a guess. The Day-One Detection Pack is a curated starting point, 63 rules chosen because they hold up in an unfamiliar environment, so a new deployment isn't running on zero.

 

All three read from the same 622-rule library, maintained centrally. That's what connects them: a rule tuned once is written as detection-as-code, immediately searchable in the Navigator, and eligible for the next version of Day-One, with nothing to keep in sync by hand. Growing from the curated 63 into the rest of the library is a deliberate process built on top of that, not a wall a team hits on day one.

 

Anomaly Detection joins them, so known threats and unknown behaviour converge on one alert and incident stream. Anomalies can also feed your alert rules, which narrows the haystack a second time.

The AI Normalizer: AI integrity regulated organisations can trust

Trust in automated detection comes from seeing how it works and knowing where your data goes. Anomaly Detection answers that by staying on your site. The AI Normalizer, a separate add-on now opening for Priority Access, shows the same commitment where a model is involved.

 

It turns unmapped log fields into usable data. Grouping, pattern-building, and typing known fields happen inside your own environment by fixed rules; a model is asked only to name the fields that step can't resolve, and only when no normaliser Guardsix already ships fits the source. Where that path runs, a bounded, redacted sample leaves your environment to Anthropic Claude, through your own subscription and key. If your organisation can't send log data off-site, the rest of Guardsix SIEM, including Anomaly Detection, is unaffected.

 

Because the model runs under your own subscription, that cost is yours to see and manage directly, never folded into a number Guardsix controls.

 

The AI Normalizer clusters unmapped syslog lines, generates a signature for each pattern, and shows exactly which fields it identified.

 

See Anomaly Detection running on your own data. Book a demo

 

Frequently asked questions

What is Anomaly Detection?

An add-on for Guardsix SIEM that learns a behavioural baseline for each user and entity and flags what deviates from it. Every anomaly carries what was observed, the baseline, why it scored, and the score.

 

When is it available?

Anomaly Detection is generally available on-premises with Guardsix SIEM 7.11. The AI Normalizer, a separate add-on, is opening for Priority Access, not general availability.

 

Does Anomaly Detection send my log data outside my environment?

No. It runs entirely inside your own deployment, with no cloud service, no external model call and no new server.

 

Do I need a new licence?

Yes. Anomaly Detection is a separately licensed add-on, off by default and not part of the base SIEM licence.

 

Does this replace my existing alert rules?

No. Alert rules catch what you know to look for, and Anomaly Detection surfaces the rest. Anomalies can also feed your alert rules.

 

Will it flood my analysts with noise?

Unusual is not always malicious, so some anomalies will be benign. The shipped rules were tuned in real design-partner environments, peer groups reduce noise, and thresholds and suppression are yours to set. The aim is a smaller, richer set to review.

 

What do I need to run it?

On-premises Guardsix SIEM 7.11, with your key identity, endpoint and network sources normalised. Active Directory, LDAP or CSV data builds the peer groups. It needs some additional memory and disk, so check the product documentation for sizing.

 

How do I join Priority Access for the AI Normalizer?

Reach out to your Guardsix contact if unmapped log sources are a live problem in your environment.