Defence AI needs controlled collaboration: what Avengers Labs teaches us

Most of the attention on the UK–Ukraine Avengers competition has gone on the data, and that’s understandable. British companies are getting access to AI training data drawn from real operations, something no lab or simulator fully reproduces.
But a lot of my working week goes on security questionnaires from defence customers, and the question I keep coming back to is a different one. How do you run a fast-moving, multinational AI programme without losing track of who can see what, who decided what, and why?
That question goes well beyond this one competition. Almost every defence AI programme now brings together government teams, primes, specialist suppliers and international partners on tight timescales. Avengers is simply the clearest current example.
What is the Avengers competition?
On 25 September 2026, the Ministry of Defence’s Rapid AI Delivery Taskforce (TF RAID) opened the first competition under the UK–Ukraine AI Partnership. Up to 12 UK companies will be selected to train and test AI models on Ukraine’s Avengers Labs platform, with a focus on drone swarms and collaborative uncrewed systems. The UK is the first international partner whose defence companies have been given access.
The competition covers four areas:
• Autonomous target recognition
• Distributed decision-making
• Adaptive mission execution when communications are limited
• Collaborative sensing across platforms
Proposals had to meet both British and Ukrainian defence requirements. Submissions closed on 22 October and selected companies were due to be told by 6 November. Phase 1 access is planned to run for around three months, from late November 2026 to late February 2027. For the companies selected, the real work starts now.
Why the access model is the interesting part
Companies don’t download the data. They bring their models to it. Selected firms work inside the secure Avengers Labs environment once Ukrainian export-control procedures are complete, and Ukraine says the data has been anonymised and delayed to reduce the risk to personnel. The competition Q&A adds that model weights licensed through the platform may only be used within the UK and Ukraine.
The scale is significant. The UK MOD says the database includes sensor information from more than six million object detections, while Ukraine describes more than five million annotated battlefield frames. They’re different measures, so they shouldn’t be added together, but either way it’s a dataset few developers could otherwise get near.
What I find most useful about this model is that governance isn’t bolted on at the end. It’s the design. The partnership declaration signed in August calls for secure data and compute pathways, joint assurance, and safeguards for data, intellectual property and export controls, while respecting each country’s sovereignty.
The information that lives outside the platform
Avengers Labs protects the training data. But the programme around it produces a great deal of information that never goes near the platform: proposals, clarification questions, due diligence, contracts, design decisions, risk reviews, test evidence and go or no-go decisions. Some of it is sensitive even though it contains no battlefield data at all.
All of that moves between UK and Ukrainian government teams, defence users, commercial specialists, technical experts and suppliers, each with their own systems and access policies. When it travels by email attachment or loosely managed shared drive, four simple questions get hard to answer:
• Which version is current?
• Who approved this change?
• What evidence supported the decision?
• Was access removed when someone’s role ended?
If the honest answer to any of those is “it’s in someone’s inbox”, that’s the gap.
Pace makes it harder. The MOD planned to go from launch to company notification in about six weeks. Governance that relies on manual chasing slows a programme like that down. Governance that gets skipped to keep up creates a different problem: months later, nobody can reconstruct why a model, a supplier or a requirement moved forward. That’s exactly what assurance reviewers will ask.
Five controls to put in place early
The fix starts with a clear information model for the whole programme, including everything that happens outside the specialist AI environment. These are the five controls I’d prioritise:
1. Classify information before it’s shared. Decide what can go into each workspace, what must stay in specialist systems, and who can approve release.
2. Use role-based access across organisations. Give suppliers and reviewers only what they need, and make access time-limited where it makes sense.
3. Keep one controlled record of requirements and decisions. Version history should show what changed, who changed it and what evidence supported the approval.
4. Link programme governance to technical assurance. Risk, issue and approval records should point to the relevant test evidence without copying protected source data.
5. Plan offboarding from day one. Removing access, retaining and disposing of records, and handing over the final record should be designed in, not left as an end-of-contract tidy-up.
Where a collaboration platform fits, and where it doesn’t
A secure collaboration environment gives government and industry a controlled place to work on approved programme information: requirements, meeting records, risk registers, contract documents, action logs, assurance evidence and approvals, with access and version history visible to the people who need them.
The boundary needs to be clear. Protected operational data stays in its authorised environment. The collaboration layer handles the documents, decisions and coordination that move the programme forward. It should sit alongside the systems used for model training, software development and testing, never become an uncontrolled copy of them.
This matters as much for suppliers as for government. A small defence technology company might need to work with MOD reviewers, an international partner and a prime contractor at the same time, without giving all of them access to its internal systems. A controlled external workspace gives everyone a shared record while each organisation keeps its own environment.
How Kahootz helps
Kahootz gives defence programmes secure workspaces for working across organisational boundaries. For a programme like this, that means controlled workspaces, granular permissions, document version control, audit trails, approval workflows, and risk and issue registers. A new project workspace can be set up in minutes and run by the programme team, without waiting on IT.
We already do this every day in UK defence. Kahootz powers Defence Share for the MOD, and we work with most of the large defence OEMs and hundreds of smaller UK defence suppliers.
To be clear about the limits: Kahootz isn’t an AI training platform, and it doesn’t replace model assurance, cyber security controls or export-control processes. Whether any platform is right for a particular set of information depends on its classification and the programme’s accreditation requirements. That’s a conversation we’re always happy to have early.
Avengers shows that speed and control don’t have to pull against each other. When access, decisions and responsibilities are designed in from the start, they reinforce each other.
Source Notes:
TF RAID Avengers: AI swarming competition (GOV.UK), 25 September 2026
Avengers AI Swarming Competition Q&A (GOV.UK), 16 October 2026
British companies to access Ukraine data to develop AI drone swarms (GOV.UK), 25 September 2026
UK–Ukraine Artificial Intelligence Partnership declaration (GOV.UK), 24 August 2026
UK becomes the first international partner to join Avengers Labs (Ministry of Defence of Ukraine), 25 September 2026
Rapid AI Delivery Taskforce (GOV.UK)


