AI agent security and AI-generated code hardening.
Before you ship an AI agent or answer a customer's security questionnaire, you need to know what the software can access, what it can change and who can stop it. We review the agents, MCP connections and AI-assisted code in an agreed scope, document the gaps and prioritise the fixes. You get technical findings your engineering team can act on and evidence your customer can assess. Corrective work is agreed separately.
What we deliver
An access register for the agents, assistants and MCP servers in scope: connected systems, data access, identities, permissions and accountable owners
A review of action controls: read and write permissions, human approval for sensitive operations, and how untrusted content is handled before tool calls
A secrets and credential review across the agreed repositories and agent configurations: storage, token permissions, exposure and revocation
An AI-assisted code review focused on authentication, access control, input validation, dependency risks and deployment settings
A logging and alerting map: recorded actions, missing signals, alert recipients and the people responsible for suspension and recovery
An evidence pack with supported findings, review limits, prioritised fixes and validation steps for agreed changes
Our approach
We start with the workflow that matters: an agent about to enter production, a customer asking for security evidence, or an AI-built application that needs a technical review. We agree the agents, repositories, integrations and environments to examine before requesting access. Then we inspect the available code, configuration and records, and carry out agreed checks of access boundaries, tool permissions and approval rules. We also examine how instructions from documents, websites or tool results are kept separate from authority to act.
The report separates observed weaknesses, controls supported by the checks performed and areas we could not verify. Each recommended fix has a reason, a priority and a way to check the result. The review and evidence pack are the starting engagement. Implementation and retesting are explicitly scoped and quoted before work begins. If the findings call for a wider codebase rescue, we explain why before proposing it; if a targeted correction is enough, that is what we recommend.
Frequently asked questions
The agents, AI assistants, MCP connections and repositories named in the agreed scope. We examine access, identities, sensitive actions, secret handling, relevant code and dependencies, and the available logs and alerts. The report records what was checked, what was found and what could not be verified. It is not an exhaustive review of your whole organisation.
Yes. We help your team answer technical questions with documented access controls, approval rules, logging, stop conditions and known gaps. The evidence describes the system as reviewed; it does not turn a missing control into a completed one. Changes that are needed are scoped before implementation, and the customer's reviewer retains the acceptance decision.
Start with the code and deployment you actually run. We examine secrets, authentication, access control, input validation, dependencies and deployment configuration within the agreed scope. These are review areas, not a claim that every AI-built application has the same defects. Local fixes can be scoped as a hardening engagement; structural problems may need a wider codebase rescue.
The initial review produces findings, priorities and supporting evidence. Fixes and retesting are included only when the quote explicitly names them. Continuous monitoring, on-call coverage and ongoing maintenance need a separate agreement; a logging review does not provide those services or guarantee a detection time.
We start with a scoping call, then provide a fixed quote and schedule for the agreed agents, repositories, integrations and checks. The quote states the access needed, deliverables and any included remediation or retesting. Additional work is agreed and quoted before it starts.
This is a scoped engineering review of applications and AI agents. It does not include a penetration test, a SOC, endpoint protection, emergency incident response or a PASSI-qualified audit. If procurement requires a qualified provider or an independent assessment, that requirement needs its own appropriately qualified engagement.
No. The engagement provides technical findings and supporting evidence, not legal compliance sign-off or certification. Your legal and compliance advisers determine which requirements apply to your organisation, product and use case. The report makes its technical scope and limitations explicit.
We agree confidentiality terms, the materials to review and the access needed before the work starts. We begin with code, configuration and documentation, using limited, temporary access where it is needed. Reports avoid reproducing secret values. Any check that could change production data must be agreed in advance.
Know what your agents can access. Decide what to fix.
Start with a scoping call about your agents or AI-built code. We define the review boundary, the evidence you need and the access required, then send a fixed quote. If the checks support your existing controls, the report says so.
