What's New · November 2026
SCYTHE 5.4
5.3 made the verdict defensible. 5.4 makes the program dependable.
Every Run
Accounted For
Held and missed runs surface for a decision
Every Host
macOS and Linux ARM64
Native clients for more of the fleet
Every Stack
QRadar and Cortex
New first-party SCYTHE Connect ingest
5.4 closes the gap between the schedule you set and the evidence you report.
★ Flagship Feature
A Schedule You Can Trust.
Proof When It Slips.
Held and missed scheduled runs now surface in Operations and the notification bell, each with a decision waiting on it.
Visibility
Seen Where You Work.
Held and missed runs appear in the Operations view and in the notification bell, ordered by due date, so the oldest gap is always at the top of the list.
Context
One Item, the Whole Story.
A held run tells you how many misses it stands for and since when. A week of skipped nightly runs reads as one item with a count, not a pile of alerts to clear one by one.
Decision
Decide, Don't Dig.
Every held run waits on an operator decision. The view also shows any runs a decision could not reach, so nothing is closed out by assumption.
Where to Find It
Open Operations or the notification bell. Each held run shows how many misses it represents, when the first one was due, and the decision it needs from you.
"If it ran, you can prove it. If it did not, you will know."
SCYTHE 5.4
What's New in 5.4
More of the Fleet.
More of the Stack.
Native clients for targeted platforms, first-party integrations with leading SIEM and XDR tools, and greater control over emulation duration, assistant workspace scope, and AI reasoning.
Connect
QRadar and Cortex XDR / XSIAM
SCYTHE Connect adds first-party ingest for IBM QRadar and Palo Alto Networks Cortex XDR / XSIAM. Cortex adds an XQL query to confirm Logged verdicts, and QRadar connects over secure TLS by default.
macOS
Native Universal macOS Client
macOS clients are now a native universal build. They carry only the modules they can actually run on a Mac, and include Python only when a test needs it.
ARM64
Linux ARM64 Clients
Build Linux ARM64 clients for Tests and pick the Linux ARM64 test client in the UI. Agents receive a client built for their own registered architecture.
Lifetime
Per-Campaign Client Lifetime
Campaign Builder adds a Client lifetime control, capped at the platform maximum. The setting travels with the client, and macOS clients shut themselves down when it ends.
MCP
Workspace-Aware MCP Sessions
An MCP session can now act in a named workspace without moving the user, and a pinned session scopes its data to that workspace. Teams and service providers keep each environment's evidence where it belongs.
AI
More Control Over AI
Set reasoning effort, point Anthropic at a custom endpoint, and use the latest Claude models. When a model declines part of a request, the builder chat now tells you what it declined and why.
Also in This Release
Quieter Installs, Cleaner Numbers
The agent installer checks for admin rights without running net session, removing an account-discovery indicator that endpoint products flagged during install. A deleted agent now stops and checks back hourly, logs how to restore the host, and no longer retries every minute. Defense Score counts integration verdicts on unassessed events, and Control Effectiveness shows Requires integration only until an integration reports. The TTX presenter view shows the participant brief and never the facilitator narrative. SSO Test Connection validates IdP certificates correctly. The AI service no longer reaches out to download models at runtime.
Spotlight Capability
You Proved What Your Stack Did.
Now the Proof Holds Up.
More connectors, steadier ingest, and correlation that stays honest when hostnames, IP addresses, and run histories get messy.
Steadier Ingest
Ingest cursors never move backward, and a full page always moves them forward. Elastic pages through every alert, and Defender paging follows its own query clock. Ingest health reports when ingest falls behind, and a paging stall reads Stale, not Failed.
Sharper Correlation
Hosts correlate by short name, IP addresses are read in full, and each detection is measured against the steps of its own device. A stray request ID is no longer counted as another test's signal.
Current, Not Cumulative
Detection analysis now scores the most recent run, not the campaign's whole history. Fix a rule, rerun the test, and the fix shows up as a fix.
Supported Connectors
Splunk · SentinelOne · CrowdStrike Falcon · Microsoft Sentinel · Microsoft Defender · Elastic Security · IBM QRadar (new) · Cortex XDR / XSIAM (new) · and more
Still True from 5.3
Detections still come back natively, every test still ends Logged, Alerted, or Blocked, and the SCYTHE MCP server still runs the whole loop from the client you already use. In 5.4, more of your fleet and your stack are in scope, and you will know when a run did not happen. Read the 5.3 release ›
5.4 by Role
Five Roles.
One Release.
The same execution data, the same permission model, and the same record of what has been validated, with 5.4 adding something every function that depends on it will notice.
Management
The validation program you funded can show it actually ran. Missed runs are counted and decided, not hidden, and Defense Score now reflects integration verdicts on events nobody has assessed yet.
Red
Emulate on macOS with a native universal client and on Linux ARM64 servers and cloud hosts. Set a client lifetime per campaign so an emulation ends when the engagement does.
Operator
Held and missed runs land in Operations and the bell with a decision attached. Agent installs stop tripping endpoint products, and deleted agents back off quietly and tell you how to restore the host.
Defender
QRadar and Cortex XDR / XSIAM join SCYTHE Connect. Ingest tells you when it is behind, correlation matches the right step on the right host, and a rule fix shows up on the next run.
AI and Automation
MCP sessions act in a named workspace without moving the user, and pinned sessions keep their data scoped to it. Tune reasoning effort, use custom Anthropic endpoints, and see exactly what a model declined and why.
Every number still links back to the tests behind it.
Now including the runs that did not happen.