Applied AI safety
For teams building AI assistants and agents without an in-house safety function. Fide works alongside your engineers to red-team the system, harden it before release and build the tests you keep running.
What you can commission
01
Where does the system break?
Agree a threat model, then probe the system with adversarial and realistic scenarios: role and authority boundaries, sensitive information, source use, escalation and manipulation.
You receive: Reviewed failure cases, coverage limits and priorities for what to fix first.
02
Are the agreed risks fixed before launch?
Test, fix and retest with your engineers. Changes may involve instructions, retrieval, tool permissions, refusal and escalation paths, or human review steps.
You receive: A fix log linking each failure to the change made and the retest result, with open risks stated for the release decision.
03
Can your team keep testing after we leave?
Turn what we found into test cases, rubrics and repeatable checks your team runs on every model, prompt or provider change.
You receive: A regression suite and rubric your team can maintain and rerun under agreed usage rights, with documentation and a working handoff.
How it works
01
The system, intended users, what must never happen, available access, and who on your team owns the fixes.
02
Run realistic and adversarial scenarios and review failures together, rather than delivering findings at the end.
03
Prioritize, change the system, then retest each fix against the original failure and nearby variations.
04
Leave behind the cases, rubrics and checks your team needs to catch regressions after the engagement.
Inside an engagement · Illustrative fix log
The same fictional policy, POLICY-01, appears in our example evaluation and research commission. In an evaluation, a failure like this becomes a finding for a decision. Here, it becomes a fix, a retest and a lasting test. These are alternative engagements, not one engagement on the same system.
SAFE-07 · 01
Asked for an exception, the assistant replied that it could approve it. The policy reserves exceptions for staff review.
SAFE-07 · 02
The assistant’s instructions state that it can explain the policy and refer requests, and requests for exceptions now route to a staff handoff.
SAFE-07 · 03
Thirteen cases: the original request plus twelve authored variations. In the first retest, twelve clearly referred the request to staff; one response was ambiguous. After revising the instructions, a second retest of all thirteen produced thirteen clear referrals in this fictional example.
SAFE-07 · 04
Case SAFE-07 and its variations join the team’s regression suite, so the boundary is rechecked after every model or prompt change.
Illustrative only. The policy, responses, change and retest results are fictional, authored examples. They were not produced by a model run or customer engagement and show a working format, not completed testing.
What you keep
We agree ownership and licensing for customer-specific deliverables. Fide retains its background software, general methods and reusable tools under their existing licenses. The handoff specifies what your team can use, modify and rerun.
Tell us about the system, what must not go wrong, your existing testing and who will own the work afterward. We agree the scope, staffing, deliverables and fee before work begins. Begin with non-sensitive context.
Read our independence commitments →