Dr Brent Coker · AI security consultant
I assess how AI applications and agents could expose sensitive information, be manipulated or take unauthorised actions—and help your team strengthen the controls around them.
Specialist focus: prompt injection defence, AI agent permissions, secure retrieval and runtime authorisation.
Identify where your AI systems are exposed, which controls need strengthening and how to verify the changes. Each assessment is scoped to your applications, data and business risks.
Assess whether malicious instructions in conversations, documents, websites or tool responses can redirect your AI system. Test safeguards and identify changes that reduce opportunities for data theft, manipulated outputs and unauthorised actions.
Examine the permissions, credentials and integrations that allow AI agents to act. Review tool access—including Model Context Protocol (MCP) connections—and identify excessive privileges, unsafe execution paths and opportunities to bypass human approval.
Assess retrieval-augmented generation (RAG) systems that connect AI to your organisation's information. Examine document permissions, customer data separation and retrieval controls to identify whether users could obtain information they are not authorised to access.
Review how proposed AI actions are checked before execution. Assess permission enforcement, approval gates and activity limits, together with the logs needed to investigate suspicious behaviour and reconstruct what happened.
Documented weaknesses, reproducible test evidence, prioritised remediation recommendations and verification of agreed fixes. The scope and testing approach are agreed at the outset.
An AI system's permissions should be enforced by the systems it connects to, independently of what the model says is allowed.
In progress
Through INTERLOCK, I am researching how independent controls can constrain AI actions before execution—including situations where a sequence of individually permitted actions could create unacceptable cumulative harm. The research focuses on high-consequence settings such as critical infrastructure.
The project examines cumulative action limits, independent authorisation and the evidence needed to explain why an action was allowed or blocked. Papers and a plain-English explainer will appear here as they are published.
I hold a PhD in electronic commerce and information systems from Victoria University of Wellington, and I have been a Senior Lecturer at the University of Melbourne since 2008, teaching digital marketing and digital business models from undergraduate to MBA level.
As founding director of Deloosh, I build and security-review LLM-powered systems in production. That practical work informs how I examine the connections between AI models, application code, permissions and business processes.
My expert-witness work includes matters involving consumer deception, fraud, privacy and data manipulation. It brings an emphasis on traceable evidence, clear reasoning and distinguishing what a system claims to do from what the evidence supports.
My research on consumer behaviour and online decision-making has been published in the Journal of Experimental Social Psychology, the Journal of Retailing and Harvard Business Review.
Research on human behaviour, digital business and online decision-making.
Tell me which AI applications or agents you use, what information they can access and what actions they can take. We can define a focused assessment around your security concerns.