
Security and Safety-Security
Research theme 03 · Led by Dr Kay Atefi
In short
On a connected physical system, a security breach is a safety event. This theme studies attacks whose end result is physical harm, and the co-assurance problem that appears when a security control and a safety argument pull in opposite directions.
Why it matters
Safety and security grew up as separate disciplines, with separate standards, teams and vocabularies. Safety assumes failures are random or systematic; security assumes an adversary actively looking for the worst case. Once a system has actuators and a network connection, those two worlds meet whether or not the organisation is set up for it.
AI makes the overlap sharper. A perception model can be attacked through its input rather than its code: a sticker on a road sign, a crafted sensor reading, poisoned training data. None of that is a software vulnerability in the usual sense, and none of it is a random failure either, so neither discipline’s standard methods cover it.
The harder problem is co-assurance. A security measure can weaken safety: an update that patches a flaw also invalidates the evidence a safety case was built on. The two arguments have to be made together.
Theme lead: Dr Kay Atefi
Assistant Professor · Programme Director, Cyber Security
Leads the security theme, working on cyber defence for AI-enabled systems and the co-assurance of safety and security.
Also contributing: Dr Funmilola Fagbola · Dr Koorosh Aslansefat
Cyber disturbance
Interference that degrades a system without taking it over — delayed messages, dropped packets, jammed channels — and how control and safety functions behave while it’s happening.
Adversarial inputs
Inputs crafted to make a model fail in a chosen direction, whether such attacks work outside the lab, and defences that hold up in a real operating environment.
Data poisoning
Attacks placed in the training pipeline rather than at runtime, including supply-chain routes through third-party datasets and pre-trained models, where the compromise happens before deployment.
Spoofing
Faked sensor and position data, including GNSS and AIS, on vehicles and vessels that act on what they’re told, and detection that doesn’t rely on trusting the same channel.
Co-assurance of safety and security
Methods for arguing both properties in one structure, so that a change made for one is checked against the other and conflicts show up during design, not during an incident.

Open questions
• How to express an adversary in a safety case, when safety analysis has no established notion of an intelligent attacker.
• Whether a security patch can be deployed to a certified system without reopening its certification, and what would make that possible.
• How far adversarial robustness results from benchmarks carry over to physical systems in the field.
Safety and security teams not talking? We work with organisations that need one argument covering both, instead of two that contradict each other.
Related themes:
05 · Robots & Autonomous Systems