Learn › Agent Harness, the big picture › Stop 7 › What a harness must do
Stop 7 of 9 · Pages
Protect: secrets, personal data and the law
The requirements a security or compliance team will ask about first.
- 30 · Keep secrets away from the model
- 31 · Protect personal data
- 32 · Keep customers apart
- 33 · Name an owner and a risk level
Keep secrets away from the model
Also called: Secrets management
Where do passwords and keys live?
Why
Anything placed in a prompt can end up in a reply. Secrets belong in a vault, used by the tool at the moment it runs.
Without it
One clever question makes the agent recite a password.
Think about
- No keys in prompts, in code or in logs
- Each tool gets only the keys it needs
- Keys that expire and are replaced regularly
In the big picture: Identity and access, in the Control family
Protect personal data
Also called: Masking, data minimisation, retention
How much personal data does the model really need?
Why
Personal data sent to a model also lands in logs, traces and memory. The law treats every one of those copies as the company's responsibility.
Without it
A data breach with penalties to match.
Think about
- What is masked before the model call, and what is restored afterwards
- Where data may be stored and processed
- How long it is kept, and how it is deleted on request
In the big picture: Guardrails, and the Governance rail
Keep customers apart
Also called: Isolation, multi-tenancy
Can one customer's data, memory or limits leak into another's?
Why
A shared platform shares its stores. Without walls, one customer's conversation can turn up in another's answer.
Without it
One customer sees another's data.
Think about
- Every stored item is labelled with its owner
- Every read is filtered by that label, in code
- Separate limits, so that one customer cannot use up another's share
In the big picture: Identity and access, and Memory and state
Name an owner and a risk level
Also called: Governance
Who answers for this agent, and how much freedom does it have?
Why
The riskier the job, the more checks and the more approval an agent needs. Somebody has to decide that, and be answerable for it.
Without it
When something goes wrong, nobody is responsible.
Think about
- One named owner for each agent
- A risk level that decides which requirements are compulsory
- A sign-off before release, and again after a large change
In the big picture: The Governance rail