Agent security and defence
Agent Security and Defence: Build, Review, Harden, Maintain
Agents read your data and act through your tools, so they need defences of their own. We design agent defence mechanisms, test them against attack, and keep them working, on top of a hardened stack from the data centre to the model.
Why agents need their own defences
A chatbot answers; an agent acts. It can read files and web pages, call tools, move data and trigger workflows. That creates risks that classic controls do not cover: instructions hidden in content an agent reads (prompt injection), tools used beyond their purpose, data leaking through outputs, and agents given more access than their job needs. Agent security treats the agent as a new kind of user, with its own identity, limits and audit trail.
Agent defence mechanisms
- Least-privilege permissions. – Each agent and each tool gets only the access its task needs, with scoped, short-lived credentials.
- Input and output guardrails. – Untrusted content is treated as data, not instructions, with injection detection and validation of what comes back out.
- Tool and action controls. – Allow-lists, rate limits and human approval for high-impact or irreversible actions.
- Sandboxing and isolation. – Code execution, browsing and file access run in contained environments with limited network egress.
- Data protection. – Separation by sensitivity, access-controlled retrieval, redaction, and secrets kept out of prompts.
- Identity and audit. – Agents have their own identities, and every action is logged so it can be traced and reviewed.
- Monitoring and response. – Behaviour monitoring, alerts on unusual activity, and a kill switch and rollback when something goes wrong.
- Red-teaming and evaluation. – Adversarial testing of prompts, tools and workflows before release and after every significant change.
Four stages
- Secure build. – Threat modelling and security requirements go into the agent design: permissions, tool boundaries, data flows and safe defaults.
- Review and testing. – Architecture and configuration review, plus red-team testing against prompt injection, tool misuse and data leakage.
- Harden. – Close the gaps found: tighten permissions, add guardrails and isolation, and fix the platform and infrastructure underneath.
- Maintain. – Monitoring, patching, re-testing on every model, tool or prompt change, and incident response so the agent stays safe after go-live.
A hardened stack underneath
- Data centre. – Physical access, segmentation of building and power systems from the IT network, and monitoring of power and cooling controls.
- Hardware. – Firmware, out-of-band management interfaces, supply-chain checks and fabric segmentation.
- Software and models. – Operating systems, containers, orchestration, model serving, model and data access, and secrets management.
- Agents. – Everything above, applied to the agent itself: permissions, tools, approvals, logging and testing.
Security testing is carried out only on systems you own or have authorised us in writing to test.
Where this fits in the stack
Common questions
What is agent security?
It is the practice of securing AI agents that read your data and act through your tools: controlling what they can access and do, defending them against manipulation such as prompt injection, and monitoring what they actually do.
What are agent defence mechanisms?
They are the controls that keep an agent inside safe limits: least-privilege permissions, input and output guardrails, tool and action allow-lists, sandboxing, human approval for high-impact steps, monitoring and a way to stop or roll back.
How is this different from conventional cybersecurity?
Conventional security protects networks, hosts and data. Agents add new risks, such as instructions hidden in content and tools that act on their own, so we add agent-specific defences on top of a hardened stack rather than instead of it.
Do you test the agents you build?
Yes. We red-team agents and their tools before release and after changes, but only on systems you own or have authorised us in writing to test.
Next steps
Related: Agent platform · LLM models and token costs · Maintenance and support · AI application layer
Want your agents reviewed and defended before they act on real data?
Agent security review
