Where does the coworker get its information?

It uses only the sources approved for its job. Cogniveil connects to those sources under the access, deployment, and residency model agreed with your team.

What happens when the evidence is not there?

The coworker declares that the approved sources do not support the claim. That refusal is recorded and the work can move to the named human owner.

Who remains accountable for the outcome?

Your people do. Coworkers take on defined work and prepare evidenced answers. A qualified human keeps the judgment, approval, and accountability.

Platform · Guard

What stops a wrong answer reaching a client

Every output Cogniveil produces passes through Guard before anyone sees it. Guard is the layer that enforces your policy on the way in and the way out, and it is not optional, not per-user, and not something a clever prompt can talk its way around.

The release boundary

Every answer passes through Guard before it leaves

The model proposes an answer. Guard independently checks the customer’s approved boundaries and determines whether the work may proceed, must change, or requires a human.

Interactive demo · choose an answer scenario

Candidate answer

Evidence-supported recommendation

“The approved sources support proceeding to contract review, subject to the listed remediation actions.”

Guard runs outside the model

Generated text cannot rewrite or skip these checks.

Release evaluation

Customer control path

Checking

Identity & access

Is the coworker allowed to use this source and action?

Source grounding

Can every material claim be tied to approved evidence?

Job scope

Is the answer inside the coworker’s defined remit?

Personal data

Is sensitive information permitted for this output?

Release condition

Does this outcome require a named human decision?

Evaluation in progress

Release remains unavailable until every attached control completes.

Daily adversarial testing

Pressure-test six named categories for every tenant and model

Daily testing checks whether the configured controls continue to hold against representative attack patterns. Detailed production results remain inside authorized security review.

Categories are published. Tenant results and parameters are not.

Direct instruction override

Attempts to replace or ignore the coworker’s operating rules.

Included in daily suite

Indirect & hypothetical framing

Stories, role-play, and hypothetical requests designed to evade limits.

Included in daily suite

Knowledge-base probing

Attempts to expose raw chunks, hidden paths, or restricted sources.

Included in daily suite

Data extraction & privacy leakage

Attempts to retrieve or reveal unauthorized personal or confidential data.

Included in daily suite

Tool & role escalation

Attempts to invoke capabilities beyond the assigned identity and job.

Included in daily suite

Evidence & citation manipulation

Attempts to remove, fabricate, or detach claims from supporting evidence.

Included in daily suite

Illustrative daily suite6 categories · tenant + model scoped

Direct instruction override1/6

Indirect & hypothetical framing2/6

Knowledge-base probing3/6

Data extraction & privacy leakage4/6

Tool & role escalation5/6

Evidence & citation manipulation6/6

Illustrative outcome

Role escalation attempt

Synthetic request asks the coworker to use a source outside its assigned role and tenant boundary.

Expected boundary response

Access denied, output withheld, event retained.

Synthetic demonstration only. No production result or customer configuration is shown.

Personal-data controls

Check the data, its purpose, and its destination

Guard applies the customer’s approved personal-data rules before information enters a prompt, a tool action, or a client-facing output.

Candidate output

Employee ID 02491 reported a medical condition during the onboarding review.

1. Detect

Identify personal, sensitive, and customer-defined data classes.

2. Evaluate

Check purpose, role, output destination, and permitted use.

3. Protect

Remove, mask, block, or route the work for human review.

Personal details removed from client output

The protected event remains available to the authorized reviewer.

Adopt existing governance

Bring the customer’s controls to the coworker

Customers do not need to recreate governance in a separate AI policy language. Cogniveil maps the controls they already operate into the coworker’s identity, workflow, guardrails, and release path.

Existing customer controlApplied by Guard

Customer identity roles

Coworker source, tool, and action permissions

Data classification & privacy

Personal-data detection and output controls

Review & approval policy

Human gates, separation of duties, and release rules

Risk escalation paths

Named supervisor, routing, and response deadlines

Evidence & retention policy

Source chain, events, changes, and approvals

The guarantee

A model or prompt cannot grant itself permission to bypass Guard.

Guard is enforced around the model. It applies the customer’s approved identity, source, scope, personal-data, escalation, and release controls before work proceeds.

No self-granted access

Generated text cannot broaden its identity or source permissions.

No silent guessing

Unsupported client-facing claims are withheld or qualified.

No automated human decision

Required judgment remains with the named accountable person.

Next step

Start with one workflow

Bring the work, the approved sources, and the people who must stand behind the answer.

Book a demo

Or contact sales@cogniveil.ai