As defence and dual-use technology races toward higher levels of autonomy, one of the hardest problems is no longer building capable systems – it’s proving they’re safe. London-headquartered Saif Autonomy, founded in September 2023, has set itself up squarely in that gap, aiming to become the safety layer that lets non-deterministic, AI-driven platforms into real-world, regulated environments.
The company will be at Apex Europe, where its technology sits neatly across several of the show’s core themes: autonomous systems, AI at the edge, and credible assurance for deployed capabilities. We talked about just what Saif Autonomy brings to that conversation.
From cockpit to code
Founder and CEO Kyle Thomas spent 16 years as an RAF pilot, finishing his military career in the Rapid Capabilities Office working in flight test and R&D on collaborative swarming programmes. The idea for Saif emerged directly from those projects.
“We were trying to develop and then deploy a capability using high levels of collaborative autonomy and non-deterministic control, and there were no guard rails,” he recalls. “You can’t certify those systems using the standards we rely on today.”
Those standards – such as DO-178C in aerospace – assume software is deterministic and its failure rates can be modelled against lines of code and hours of operation. Physical AI systems, by design, don’t behave that way.
“You don’t know what the output is going to be – that’s the point,” Thomas says. “It’s very good at goal-based autonomy, but there was nothing that let regulators or operators trust it.”
A “trust layer” for non-deterministic systems
Saif Autonomy’s answer is a high-integrity “trust layer” that sits between any untrusted input, whether that’s a human, a deterministic system, or a physical AI model, and the platform’s controllers, sensors and effectors.
At runtime, this software intercepts the command stream, predicts forward faster than real time, and checks intended behaviour against configurable rule sets:
- legal constraints and rules of engagement
- safety envelopes
- containment and geocontainment rules
- sensor and effector performance thresholds
If the system is about to breach those rules, Saif’s software intervenes, adjusting or vetoing commands while keeping the platform, as far as possible, “on task”.
“We create a bubble within which you can use non-deterministic AI,” Thomas explains. “You can do all the things you want inside that bubble, but you’re never going to leave it – and you’re never going to do something you absolutely shouldn’t.”
Crucially, this can run at the edge, on-board aircraft, ground vehicles, maritime systems or space platforms, without relying on data links – making it relevant for denied and contested environments. The same architecture can also be deployed centrally, with a lightweight safety layer at the edge for “hard stop” actions.
Top-down certification, not bottom-up patches
Where many newer entrants aim for easier commercial use cases first, Saif has gone directly at the hardest end of the assurance spectrum, targeting the highest levels of aerospace safety certification and then “waterfalling” down into other domains.
“Competitors are going for the easy wins first,” Thomas says. “But if you start bottom-up, you end up having to redo everything for safety-critical certification. There’s a body of evidence and time you simply can’t buy, no matter how much funding you have.”
The company has patented its architecture and is pursuing the highest integrity levels of certification. That approach has resonated more quickly with large defence primes than with smaller players:
- BAE Systems, Thales, Applied Intuition and others are already customers.
- UK defence users and organisations such as DSTL have trialled and validated the technology.
- In the US, interest from bodies like the Air Force Research Lab has helped unlock early investment.
Thomas admits he initially expected Saif’s early adopters to be smaller firms without in-house safety-critical expertise. In reality, big primes have moved first, seeking an assured layer they can embed into their platforms as a differentiator against lower-cost rivals.
Dual-use ambitions, defence-first validation
Although firmly dual-use, Saif Autonomy has so far been defence-heavy. The company operates a 100+ acre test site in Oxfordshire with runway and flying testbeds, allowing experimentation in unsegregated airspace, and has a footprint in Sunnyvale and Austin to support the US market.
The aim now is to take a defence-validated capability into:
- subsurface and off-road autonomy
- maritime systems
- commercial physical AI in sectors like mining, smart infrastructure and robotics
A complementary product, Prism, analyses the rich safety and behaviour data generated by Saif’s core runtime. That data can be used to retrain models, refine safeguards and, in commercial environments, potentially be anonymised and shared to lift safety baselines across entire sectors.
Funding, focus and the coming shakeout
Saif Autonomy is venture-backed with a mix of UK and US capital, but Thomas is wary of the “growth at all costs” trap:
“VCs often want you to grow way beyond your means early so you’re addicted to raising,” he notes. “We’ve tried to be clever – grow fast, but only hire when we need to service demand.”
Unlike many early-stage defence tech firms, Saif has generated significant revenue from day one, blending R&D and delivery contracts and using that traction to validate its approach.
At the same time, Thomas is clear-eyed about the broader market: a flood of VC money into physical AI, a proliferation of drone and autonomy companies, and a likely correction that will leave only a handful of “category winners” standing. His view is that safety and assurance will be decisive.
What to look for at Apex Europe
At Apex Europe, expect Saif Autonomy to position itself less as another autonomy stack and more as the enabling layer that lets everyone else’s autonomy actually deploy – in airspace, on land, at sea and in orbit.
In a landscape where physical AI is advancing faster than regulation and assurance, Saif is betting that the real leverage lies not in adding more intelligence, but in making that intelligence certifiable, controllable and legally usable. For defence customers and dual-use innovators alike, that may prove to be the difference between an impressive demo and a fielded capability.
