AI and agent security
Security testing for AI systems and agentsTest what they can reach and what they can do.
An AI system is more than a model. It has data sources, permissions and tools. We test the whole setup, because that is where damage happens.
What we test
What the assessment covers
- 01
Prompt injection
Can text in a document, web page or message change what the system does?
- 02
Data exposure
Can a user pull out information they should not see, including from retrieval sources?
- 03
Excessive permissions
Does the agent have more access to tools and data than its job needs?
- 04
Unsafe actions
Can a manipulated agent send, delete or change things without a person approving?
- 05
Surrounding controls
Logging, rate limits, approval steps and secrets handling around the model.
- 06
Third party components
Models, plugins and services the system relies on.
Scope
What we assess
- Assistants, chatbots and internal knowledge tools
- Agents with access to tools and company data
- Retrieval and document pipelines
- Systems built by your team, by us or by another vendor
Exclusions
What stays out unless agreed
- Attacks on systems you do not own or control
- Social engineering of your staff unless agreed in writing
- Denial of service testing unless agreed in writing
Method
How findings are produced
- 1
Scope
We agree what is tested, the access we need and what is out of bounds.
- 2
Test
We test the running system and review code, configuration and dependencies.
- 3
Validate
We confirm each finding is real and rank it by risk.
- 4
Fix and retest
We fix what we can with you, then check every fix.
The report states who performed the review and what it covered. An assessment of another vendor’s system is described as such. No assessment can promise complete protection.
