Security

How do you verify AI without trusting what it says?

Every model response, reviewer output and retrieved source is treated as untrusted input. Trust is earned through isolation, validation, evidence and provenance.

Follow a threat through the defenses.

Choose a scenario: the diagram shows where Recherche.fr must break the trust chain.

01Evidence injection
02Persuasive output
03Workspace contamination
04Provenance loss
Trust chain

See exactly where trust must be broken.

The selected scenario above moves the defensive boundary. The principle remains: no AI output or retrieved evidence becomes trusted instruction by default.

Hostile input

Boundary

ControlIsolation / validation

Content remains untrusted during processing.

OutputAuditable trace

The verdict keeps its link to evidence and controls.

Explanatory threat modelClaims remain limited to documented controls

The threat model starts with untrusted content.

01

Untrusted-input separation

Provider responses, reviewer responses and retrieved pages are handled as untrusted content rather than instructions.

02

Workspace isolation

Research sessions and workspace data are kept separated so one workspace does not become implicit context for another.

03

Server-side validation

Inputs are validated server-side; public forms use bounded payloads, honeypots and rate controls.

04

Provenance

Outputs, claims, evidence and verdicts are designed to remain linked so the final synthesis can be audited.

05

Content hardening

CSP, frame restrictions, content-type protections, restrictive robots handling and WAF controls are part of the current public-site baseline.

06

Privacy-aware design

Redaction, controlled access and privacy-safe evaluation data are part of the broader architecture and ongoing R&D.

Verification-specific threat model

The verifier can itself be attacked, misled or contaminated.

01

Retrieved-content injection

A page used as evidence may contain instructions intended to manipulate the verifier. Retrieved content must remain data, not authority over the system.

02

Persuasive model output

A confident or well-written answer can bias a reviewer or Judge. Evidence quality must outweigh style and confidence.

03

Cross-espace de travail contamination

Data from one workspace must not become hidden context for another research session.

04

Provenance loss

If the chain from claim to source to verdict is broken, the final synthesis becomes harder to audit and easier to over-trust.

What is implemented today — and what is not yet claimed.

Current baseline

  • Content Security Policy and restrictive response headers
  • WAF rules for suspicious methods and scanner paths
  • Rate controls on public contact endpoints
  • Server-side form validation
  • Workspace-oriented application architecture
  • No public claim of certification or audited compliance

Ongoing hardening / R&D

  • Expanded provenance services
  • Privacy-safe evaluation datasets
  • Adversarial testing of Judge and verification flows
  • Scale, resilience and model-governance controls
  • Security validation in representative industrial pilots

Security claims stop where the evidence stops.

Recherche.fr distinguishes implemented controls from planned hardening and research work. It does not turn a security roadmap into a certification claim, and it does not claim an audit that has not occurred.

Threat-model walkthrough

A threat crosses several boundaries before a verdict.

Select a scenario: the system shows the trust boundary, applied control, remaining uncertainty and retained trace.

Threat model · illustrative
Threat
Threat
Trust boundary
Control
Remaining uncertainty
Trace retained