IN BRIEF
AI security covers measures to protect AI applications, models and agents against misuse, manipulation and unintended data leakage. Language models introduce additional attack paths, such as prompt injection. When agents access data and tools, familiar security problems such as overly broad permissions can also have greater consequences.
When to involve us.
When AI assistants access internal documents, agents operate tools and interfaces, or business units build their own AI applications, attack surfaces and permissions should be reviewed before these solutions go live.
We identify which AI applications and agents exist, which data and tools they access, and where models and components come from. On that basis, we define least-privilege permissions, safeguards against prompt injection and data leakage, and approval points for critical actions. Before go-live, we use red-teaming tests to assess the measures against selected realistic attack scenarios and document the weaknesses and limitations identified.
What the work produces.
We agree on the scope around your goals, starting point and available capacity. The following elements illustrate what an engagement can include.
WHAT WE ADDRESS
- Inventory and threat analysis for AI applications
- Permission model for agents and tool access
- Safeguards against prompt injection and data leakage
- Red teaming and model supply-chain review
POTENTIAL OUTPUTS
An assessment of AI-specific risks, a permissions model and documented test findings on safeguards. These identify required improvements and unresolved risks to inform the decision on production use.
A focused starting point.
A list of the AI applications in use or planned, with their data sources and connected systems. Information security, data protection and the responsible business units should be involved together.
The question first. Then the scope.
The first conversation clarifies the objective, deliverables and people involved. We agree whether the engagement covers design, pilot support or technical implementation, who takes on each task and how the deliverables will be assessed.
Discuss a starting point by emailWhy aren't our existing security controls enough for AI agents?
Many existing controls remain important, such as identity and access management, logging and network segmentation. What is new is that a language model cannot reliably tell instructions apart from data. A crafted document, email or web page can lead an agent to disclose confidential information or take unintended actions. The more tools and permissions an agent has, the greater the potential damage. That is why agents need narrowly scoped rights, approvals for critical steps and monitoring that detects unusual behaviour. Targeted attack tests before go-live show where the limits of the safeguards lie.
Last updated: