Defined scenarios
Simulated target scenarios and rules of engagement establish the scope of each evaluation.
CRUCIBLE
Bounded evaluation for AI agents in cyber operations.
CRUCIBLE evaluates and benchmarks AI agents for defined scenarios and policies. It produces bounded evidence for an evaluated release—not an authorization decision.
- Defined scenarios
- Distinct evaluation modes
- Bounded evidence
What we do
Evaluation evidence, kept in scope.
ATT&CK-grounded measurement
Scored trajectories use MITRE ATT&CK vocabulary within stated scenario and policy boundaries.
Separate evaluation modes
Agent-alone benchmarks and human-plus-agent evaluations are distinct measurements; results are not interchangeable.
Evidence for defined decisions
A release-, scenario-, and policy-specific result can inform a sponsor or customer decision; it does not replace authorization.
Delivery boundary
Designed for customer-led enclave integration.
Delivery is designed to support customer-led integration into a customer-controlled IL7/TS enclave. The customer, sponsor, and AO own the boundary, classification, dependencies, IdP, PKI, KMS, SIEM, storage, network, CDS, host, operations, and authorization decisions. The public-AWS rehearsal and pre-decisional package are not a customer deployment, an authorization, or proof of operational control effectiveness.
Documentation
Read the canonical guidance.
Evaluation methodology
Scope, measurement modes, reproducibility, and the current nominal-calibration status.
Deployment guidance
Use the release-matched guide for customer-led deployment and integration.
Evidence & authorization boundary
Customer-furnished controls and authorization remain outside this product.
Acquisition
Commercial context and bounded evaluation options.
Role guides
Canonical console and role guidance lives with the integration documentation.
Contact
Get in touch
For evaluation, customer-led integration, and acquisition inquiries: