Menu
Red teaming · Practical security · Advanced research
AI security, from testing
to real-world resilience.
Expose weaknesses. Put safeguards into practice. Investigate what comes next. Deep Cyber brings together adversarial testing, practical AI security and applied research.
Led by John Sotiropoulos · UK Government implementation guide author · OWASP leader
See our work with Lloyds Banking GroupFeatured work · Financial services
See how we helped Lloyds Banking Group red team customer-facing agentic AI.
Read the story in Infosecurity MagazineJohn Sotiropoulos helped Lloyds Banking Group apply the OWASP Top 10 for Agentic Applications to red teaming.
Explore the collaboration and the bank’s practical approach to securing AI agents—from controlling tool access and detecting agent hijacking to embedding security across the development lifecycle.
As reported by Infosecurity Magazine · 5 June 2026How we help
Three ways to strengthen
your AI security.
From finding exploitable weaknesses to implementing controls and resolving difficult research questions.
01 / Test
Red Teaming &
AI-Assisted Pentesting
Adversarial testing of AI applications and agents, alongside AI-assisted penetration testing of wider systems. Expert judgement and targeted automation turn weaknesses into validated findings and clear remediation priorities.
What an engagement covers
- Prompt injection, data exposure and agent tool misuse
- Application and infrastructure penetration testing
- Evidence, remediation priorities and retesting
02 / Operationalise
AI Security
Made Simple
Put the UK AI Cyber Security Code of Practice to work. Translate security principles into practical safeguards, clear ownership and evidence that controls operate—supported by the author of its implementation guide.
What an engagement covers
- AI inventory, threat assessment and prioritised gaps
- Controls, accountable owners and an implementation roadmap
- Monitoring, incident response and assurance evidence
03 / Investigate
Advanced
Research
Applied research into AI forensics, resilience and sovereignty. Reconstruct agent behaviour, evaluate failure and recovery, and understand control over models, data and infrastructure—with evidence to inform your next decision.
What an engagement covers
- AI and agent forensics: behaviour, evidence and incident reconstruction
- Resilience: adversarial evaluation, failure and recovery
- Sovereignty: dependencies, architectures and technical control
The expertise behind Deep Cyber
Helping shape the guidance.
Putting it into practice.
Deep Cyber is led by John Sotiropoulos, commissioned by the UK Government to author the implementation guide for its AI Cyber Security Code of Practice.
John chairs the OWASP Top 10 for Agentic Applications, co-leads the Agentic Security Initiative and is the author of Adversarial AI Attacks, Mitigations, and Defense Strategies.
That perspective connects security testing, implementation and research—helping your teams turn complex AI risks into practical action.
AI Cyber Security Code of PracticeRead the guide and author attribution ↗Security community leadershipOWASP Agentic SecurityView leadership roles and contributions ↗Research · Coauthored preprint · September 2026Redefining AI Security
Incident Reporting for AgentsExplore the research ↗
Operationalising AI security
From the UK Code
to everyday practice.
Make AI security part of how you design, deploy and run systems. Start with your risks and responsibilities, then build the controls and evidence your organisation needs.
Discuss implementation supportGrounded in the UK AI Cyber Security Code of Practice and ETSI EN 304 223.
- Understand your exposureIdentify AI systems, relevant threats and the gaps that matter most.
- Put controls to workAgree priorities and ownership. Implement safeguards across architecture, suppliers and delivery.
- Operate and demonstrateEstablish monitoring, response and recovery processes, with evidence that your controls work.
From the Deep Cyber blog
Ideas grounded in practice.
Sovereignty
Can Europe Use Open Weight Models to Accelerate Sovereign AI?
Explore the role of open-weight models in Europe’s sovereign AI strategy.
Agentic security
When Agents Go Rogue
Why rogue agency is an emerging systemic risk for the enterprise.
Practical implementation
Securing AI Infrastructure
A practical, risk-tiered playbook aligned to ETSI EN 304 223 and the UK Code.
What do you need to secure?
An AI system ready for testing. A security programme to put into practice. A research question that needs evidence. Let’s define the next step.
Talk to Deep Cyber