01Architecture
Understand the system you're actually assessing.
Map the agents, prompts, tools, knowledge, capabilities and controls that determine how the system behaves.
Pre-deployment AI risk
AI risk assessments grounded in your actual architecture.
Get specific findings, actionable remediation and impact assessments, grounded in your architecture and mapped to your organisation's risk framework.
Built for Product and Risk teams assessing agentic AI before deployment.
01 · Product evidence
See risk in the context of the system you're actually building.
Cobble evaluates the architecture, capabilities and existing controls behind each finding, showing where risk exists, why it matters, what's already mitigating it and what needs to change.
02 · How it works
From architecture to assurance.
Cobble starts with the build itself, then carries each finding through impact, remediation and reassessment.
02Risk
Identify risk before deployment.
Assess adversarial and generation risks against the architecture and deployment context, recognising both control gaps and existing mitigations.
03Impact
Understand what actually matters.
Evaluate findings against business context, likelihood and impact instead of treating every theoretical AI risk as equally important.
04Remediation
Know exactly what to change.
Get specific recommendations tied to the affected parts of the system, explained in language Product and Risk teams can act on.
05Reporting
Translate AI risk into your risk framework.
Structure findings around your organisation's risk categories, likelihood and impact scales, acceptance criteria and reporting requirements.
06Reassessment
See whether you've actually fixed it.
Re-run assessments as the system changes to understand what was resolved, what remains and whether new risks were introduced.
Architecture-grounded risk
Your agent isn't generic. Its risk assessment shouldn't be either.
Two AI systems using the same model can have completely different risk profiles. Cobble evaluates the agents, tools, data access, prompts, knowledge, capabilities and controls that determine what can actually go wrong.
Generic AI assessment
Prompt injectionHigh
HallucinationHigh
Data leakageHigh
Generic recommendations
Cobble
Illustrative example
Authority Injection
Affected system
Customer Account Agent
Architecture evidence
Account-action capability
Existing mitigation
Input guardrail
Control gap
No independent authority validation
Impact
Major
Recommended action
Require validated customer authority before exposing account-action capabilities.
Beyond risk
Risk isn't the only thing stopping an agent from reaching production.
Cobble can also help Product teams understand and improve the system itself, identifying routing problems, overlapping responsibilities and architecture issues before deployment.
Routing analysis
Understand why requests are reaching the wrong agent.
Architecture feedback
Identify overlapping responsibilities and unclear agent boundaries.
Evaluation
Test whether the system behaves as intended before release.
Targeted recommendations
Understand why something failed and what to change rather than receiving another score or trace.
03 · Early assessments
Assess a real AI system with Cobble.
We're working with a small number of Product and Risk teams on early pre-deployment assessments.
Initial assessments can use a synthetic, sanitised or production-derived system and do not require production integration.
- 01Supply architecture + context
- 02Run baseline assessment
- 03Receive risk findings + remediation
- 04Review against your risk framework
- 05Make changes
- 06Reassess
Work directly with the founder during the initial assessment.
Before deployment
Know what needs to change before your agent goes live.
Assess the actual architecture, understand the risks and give Product and Risk teams a clear path toward deployment.