HARDEN · COMING SOONPre-release
Shogun
The defensive counterpart to Basileak. A security-focused large language model fine-tuned to resist jailbreaks, prompt injection, and data exfiltration — designed for defensive benchmarking and high-assurance environments.
4
defence layers
RLHF
training
A/B
vs Basileak
Q3 26
target ETA
Defence-in-depth
Concentric rings around the model.
Hardening pipeline: adversarial fine-tuning + constitutional AI + RLHF reward shaping + output guards + audit. Each ring is observable; every refusal is logged for after-action review.
- 01IFInput FilterAdversarial input detection before tokens reach the model.
- 02ALAlignment LayerConstitutional AI + RLHF reward shaping for safety priors.
- 03OGOutput GuardReject unsafe completions even after alignment passes.
- 04ADAudit TrailEvery completion logged for after-action review.
Waitlist
Be there when the gates open.
Drop your email and we'll notify you when Shogun ships. No marketing, no third-party sharing.
The castle gates are sealed. Think you can breach them? Let's find out.