AI Safety & Security · Technical
Realistic, contained ranges for evaluations, rebuilt exactly for every run.
Coldlight builds each range from a defined specification. It runs contained: nothing inside reaches the internet unless you decide it should. Rebuild it exactly, as often as you need. Every run leaves a record.
An evaluation is only as realistic as the environment it runs in.
Each range can include hosts, applications, synthetic users and realistic activity, at the fidelity each test needs.
Customer-selected models, agents and tools run inside it, with the access each test gives them.
What you can test
- 01Give agents real tools and real permissions inside a contained network.
- 02Measure a model’s cyber capabilities against realistic, multi-host networks.
- 03Reset to the same starting state for every run, and compare the results.
- 04Test safeguards and monitors: see what they catch, and what they miss.
- 05Run attacker, defender and user activity in the same range.
- 06Review every run: what the model or agent did, and what changed.
What a range can include
01
Hosts and networks
Workstations, servers and the networks between them, with real operating systems where the test needs them.
- Workstations
- Servers
- Networks
02
Applications and cloud
Business applications and cloud services for agents to work in.
- Business applications
- Cloud services
03
Users and activity
Synthetic users and realistic activity, so the range behaves like a working organization.
- Synthetic users
- Realistic activity
04
Security tools
Monitoring and detection tools, and the safeguards under test.
- Monitoring
- Detection
- Safeguards
05
Industrial systems
Industrial control systems, where the test needs them.
- Control systems
06
Models and agents
The models and agents under evaluation, with the tools and access each test gives them.
- Models
- Agents
- Tools
Fidelity
Level 1
Containers
Cost-efficient representations of hosts, services and networks.
Level 2
Virtual machines
The real operating systems and software that your systems run.
Level 3
Appliances
Real or virtual appliances, where the test depends on them.
Synthetic users, scripted or live adversary activity, and defensive tools or agents can run in the same environment.
Third-party software can run under licences you already hold, or we can arrange it as a cost of the engagement.
The rules
EU AI Act, Article 55
Providers of general-purpose AI models with systemic risk must evaluate their models with standardised, state-of-the-art protocols and tools, including documented adversarial testing, to identify and mitigate systemic risks.
EU AI Act, Article 113
Chapter V, on general-purpose AI models, applies from 2 August 2025.
Every environment
Contained by default
What runs inside an environment stays inside it unless you decide otherwise.
Rebuilt exactly
Environments are defined as code. Reset one, change one thing and run the test again.
A record of every run
Every run leaves a record you can review. Compare results across runs and configurations.
The fidelity each test needs
Use lightweight representations where they are enough, and real software and appliances where the test needs them.
Technology-neutral
Open-source, commercial, proprietary and customer-built technology can all be included.
Open formats
Environment definitions use an open, portable specification, so your work is not locked to us.
Where it runs
Hosted
Coldlight as a service. Hosted environments can run in Canada and other jurisdictions.
On-premises
Coldlight inside your own boundary, on infrastructure you control.
Air-gapped
Coldlight on networks with no outside connection.
We can also operate Coldlight for you, and help design environments, deliver exercises and support evaluations.
