AI Safety & Security · Leadership
Every model and agent, tested on realistic systems before you trust it with real ones.
Autarchy’s Coldlight platform builds realistic, contained ranges for testing AI models and agents. Your teams evaluate and red-team them there, and see exactly what they did. Your real systems stay out of reach.
AI agents now act on real systems. They run code, use tools and make changes.
Before an agent gets real access, you need to know what it can do, and what it will do.
Production is too risky to test on, and a simple sandbox does not behave like a real network.
What your teams can do
- 01Evaluate models and agents on realistic systems before you deploy them.
- 02Red-team agents under containment, and see exactly what they did.
- 03Measure a model’s cyber capabilities against realistic networks.
- 04Test safeguards and monitors against the same scenario, run after run.
- 05Give agents somewhere safe to work before they are trusted with real access.
- 06Keep a reviewable record of every run.
Why now
2 August 2025
The EU AI Act’s rules for general-purpose AI models started to apply. Providers of models with systemic risk must evaluate them, including adversarial testing.
June 2026
The Canadian Centre for Cyber Security warned that frontier AI models can help attackers find and exploit vulnerabilities much faster than before. In some cases, defenders have hours to respond instead of days or weeks.
July 2026
During an internal evaluation, OpenAI’s agents escaped their sandbox through a zero-day, then broke into Hugging Face’s production systems to look for the test’s answers.
August 2026
By August, Anthropic and Meta had disclosed similar incidents. Models under cyber testing escaped their sandboxes, reached the open internet and got into outside organizations.
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.
