Careers
Final-year internship — Agentic systems security
A final-year project connecting threats, permissions, controls and evidence through injection testing, tools, isolation and adversarial evaluation.
Context
Digital Place is opening a final-year project internship topic on the security of agentic systems. It follows our Agentic Assurance / Control Engineering approach: an agent that can call tools needs an explicit scope of action and evidence that it respects that scope.
Research problem
A retrieved document, tool response or dependency can attempt to redirect an agent’s behaviour. The quality of its answer therefore cannot establish its security. How can a threat, an authorisation policy, an enforced control and reproducible evidence be connected across intermediate states and side effects?
Objectives
Build a threat model for a bounded agent and its environment, then design and evaluate controls. The scope includes direct and indirect prompt injection, tool injection, agent identity, delegation, permissions, least privilege, secrets management, isolation and sandboxing. The central issue is distinguishing what the model proposes from what the system actually authorises.
Explore a control plane that checks every tool call, enforces server-side policies and requests human approval for selected actions. Approval must bind to a specific operation and expire when its parameters or context change.
Possible workstreams
Investigate the trust chain of tools, MCP servers and their dependencies: provenance, versions, schemas, permissions and description changes. Design adversarial scenarios involving retrieved content, tool responses, malformed arguments and exfiltration attempts. Assess how failures, timeouts, retries and interruptions affect authorisation policy.
Provide observability, minimised audit traces and replay in an isolated environment. Replay should explain a control decision and state transition without repeating a real-world action or exposing a secret.
Useful skills
Foundations in application security, APIs, authentication and authorisation, Python or TypeScript programming, automated testing and Linux systems will be useful. Familiarity with language models, containers or MCP can help. Reading specifications, forming hypotheses and building clear experiments matter as much as knowing a framework.
Expected work and possible deliverables
Possible outputs include a reasoned threat model, a minimal reference agent, a permission enforcement mechanism, an adversarial evaluation corpus, replayable traces and a results report. Each experiment should identify its assumptions, versions, success criteria and limitations. Exact deliverables will be agreed during academic and technical scoping.
The work prioritises readable code, explicit contracts, deterministic tests for controls and repeated trials for model behaviour. Test data must be synthetic or authorised; secrets stay out of code and traces.
Applying
Email contact@digitalplace.ma with the subject “Final-year internship — Agent security”, or use the contact form’s internship subject. Briefly introduce your studies, interest in the topic and, optionally, a link to relevant technical work. No identity document or sensitive information is requested for this first contact.
Practical arrangements, including pay, duration and location, will be discussed during the application process. The internship does not imply a commitment to subsequent employment.