An AI-driven Red Cell for UK resilience exercises
Feliks.ai is an AI-driven Red Cell that plays an adversary's hybrid-warfare apparatus in the TTX. It plans a Red campaign from a human-authored strategic directive, declares Red's moves each turn, and drafts documents that a White Cell would otherwise write by hand.
It is guided by Russian and other CRINK strategic culture, doctrine and risk tolerance: it is a machine, relying only on Russian/CRINK doctrine and Western professional analysis. It emulates a three-star intelligence commander running ten hybrid warfare agencies, ranging from disinformation to cyber to sabotage and intelligence collection.
At STARTEX, it not only has an Operational Order (OPORD) and an intelligence assessment; it has an Order of Battle (ORBAT) including assets positioned in, or aimed at, your LRF's region. It executes its mission without sentiment.
Feliks.ai is a combination of agentic AI model, harness and loop.
The latest version of Claude. It works from structured doctrine files, an SOP and task-specific skills loaded for each game.
A collection of skills and instructions to Claude developed by academics at the University of Exeter, making the game auditable and repeatable. These generate:
A mixture of automated tasks and human stop-checks and inputs, programmed by the Umpire, and illustrated below.
Once the TTX is over, the Umpires have the option to retain all data (for example CoAs, RFIs, Red OPORDs and reasoning) for analysis, generating scripted scenarios or CoA analysis.
Feliks.ai works from a curated open-source evidence base.
The average time to generate an OPORD, ORBAT or Red CoA is under 10 minutes. AI also allows on-the-fly generation of immersive content: if an incident happens unplanned in Turn 1, there can be a video report of it in Turn 2.
Since all information is gathered by AI from OSINT only, participants can use a detailed common picture with no risk of classified material leaking into the public domain. This makes it ideal for a mixture of cleared and non-cleared professionals. Players can either command their real-world region or work in a generic sandbox region named Vallen.
Before each TTX, Feliks.ai performs a scheduled scan of the global media, updating a file of all attributed and likely Russian/CRINK hybrid actions, adding new units to the ORBAT where discovered, and altering its own pattern of behaviour. TTX participants fight the most up-to-date and realistic version of Red hybrid efforts available from OSINT.
Feliks holds in its context folder a 6,000-word document on Russian/CRINK hybrid doctrine, condensed from 27 PDFs of think tank, public intelligence and analyst reports. This machine-produced, human-amended playbook generates a wide range of potential attack scenarios, and updates automatically after each OSINT scan.
Since the exercise is unclassified, all data input is via laptops in an online form (click here to see example). Every CoA and RFI is captured, and Feliks generates routine and bespoke analysis of the developing situation.
In the course of developing this solution, the Claude model has moved from v4.0 to 5.5, improving the speed and quality of Red responses. In parallel the harness has evolved through trial and error, and now relies on skills that can be improved or augmented.
A five-cell TTX needs six laptops on the same network (one for the Director or presenter) plus one laptop or desktop to run Claude Cowork. Because Feliks.ai runs separately from the Blue network, the TTX can be run by a mobile team using the client organisation’s networked IT, such as a university or local council.
Red’s doctrinal tendency to improvise attack surfaces from Blue moves is the most surprising departure from the standard, semi-scripted TTX. It can be dialled down or up on request, allowing Directors and clients to stage a largely unscripted TTX, defining only the broad premise rather than pre-preparing branching paths and content.
In theory, the number of Blue cells could be large, and include remote cells and even AI cells played by another instance or model.
Under close human control, the agent will self-improve if allowed.