Security

Security & Trust

The Fenrir Trust Framework is the set of engineering and policy commitments that govern how FENRIR is designed and deployed. We invite scrutiny before deployment — not after.

Air-gapped by default

FENRIR runs entirely on-premise with local AI models. No data leaves the customer's infrastructure. No cloud dependency. No outbound network as an exfiltration vector.

Natively suitable for SCIFs, tactical environments without connectivity, and environments requiring IL-4/5/6 classification levels.

Other AI tools in this category typically deploy via cloud APIs to foreign LLM providers. FENRIR keeps data physically inside the customer's walls.

Deployment options

Air-gapped on-prem — Default. Local AI models, zero external dependencies.
CJIS-compliant cloud — AWS GovCloud or Azure Government where required.
Hybrid — Combined topologies for mixed-classification environments.

Architectural ban on contamination

Police interrogation and military HUMINT share the same fundamental risk: AI suggestion can contaminate the interview itself. The legal frames differ, but the engineering response is identical.

Permitted

  • Memory and coverage assistance
  • Recall of verified case-file information
  • Framework adherence scoring
  • Bias pattern detection

Prohibited — non-configurable

  • Suggested questions or interrogation tactics
  • Deception or credibility scoring
  • Motive or character inference
  • Emotion inference

The interviewer authors every question. FENRIR supports the interviewer's judgment; it does not substitute for it. These prohibitions are architectural — they cannot be disabled by deployers.

The Trust Framework

01

Grounded Outputs

The system generates content only from verified source material, with every output traceable to its origin in the transcript. No hallucinated conclusions — every finding links back to a specific moment in the interview.

02

Calibrated Uncertainty

The system communicates its confidence and abstains rather than confabulates. When FENRIR is uncertain, it says so — rather than generating plausible-sounding analysis that cannot be verified.

03

Layered Human Oversight

Human verification is required at multiple levels. Core safeguards are non-disableable by deployers — addressing a documented failure mode in current AI policing tools where agencies disabled hallucination safeguards.

04

Provenance & Attribution

AI and human contributions are distinguishable at the sentence level. Audit logs are immutable and disclosure-ready for Brady obligations, court-martial discovery, and EU AI Act Art. 13.

05

Demographic & Linguistic Fairness

Performance is measured and published across the customer's operating language and dialect set. Addressing documented ASR disparities by race and accent is a precondition for both civil-rights compliance and intelligence quality.

06

Security & Sovereignty

Air-gapped on-prem with local AI models is the default; cloud is the option. CJIS-compliant cloud, IL-4/5 cloud, and hybrid topologies are available where required. No data leaves the customer's infrastructure.

07

Accountability Ecosystem

External validation through ISO/IEC 42001 certification, independent audits, expert-witness readiness, mandatory AI literacy training, and a whistleblower channel.

Ready to close the intelligence gap?

Get in touch to see how FENRIR can transform your interview analysis.