Security, for the CISO’s first questions.
Where YMUX runs, what reaches the model, what each agent may touch, and what we do not have yet.
Deployment
YMUX runs on the customer’s servers (on-prem or private cloud). Source code does not leave the customer’s environment. Model calls go to Amazon Bedrock in the customer’s AWS account.
Your servers
YMUXworkflow engine and agents
Your repositorycode stays here
Your databaseaudit records
Task, relevant files, plan
Your AWS account and region
Amazon BedrockClaude (Anthropic)
AI models
- Models run through Amazon Bedrock, in the customer’s own AWS account and region.
- Today’s model: Claude (Anthropic), served by Bedrock.
- Amazon Bedrock does not store prompts or outputs, and does not use them to train models. Source: AWS Bedrock data protection.
- Sent to the model: the task, the relevant files, the plan.
- Not sent: secret files and anything outside the agent’s file list.
- Option to use a different model or runner (LangGraph runner supported).
Agent permissions
- Each agent role has a written list of files it may read and change.
- Blocked during the run: pushes, history rewrites, reading secret files, actions outside the role’s scope.
- After each run, YMUX checks what the agent actually touched against its list.
Audit record
- What: request and confirmed goal, who approved each step and when, which checks ran and their results, which model ran with tokens and cost, files changed and each agent’s file limits. See a sample record.
- Where: the customer’s database.
- Retention: set by the customer.
DORA Art. 9(4)(e)
Every ICT change must be recorded, tested, assessed and approved.
| DORA requires | YMUX step |
|---|---|
| Recorded | Request, goal, plan and full audit record per change |
| Tested | Tests written before code; automatic checks after each step; past checks re-run before merge |
| Assessed | Review agent compares code to plan; it can block, not approve |
| Approved | Named human approvals at goal, plan and sign-off |
Certifications
No SOC 2 or ISO 27001 yet.
Both are planned.
Contact
Security questions and disclosure: [email protected]