AI Safety
Last updated 12 June 2026
Our approach
We build AI systems that operate inside a client's existing controls, not around them-every agentic workflow we ship has a defined boundary of what it is permitted to do, and an audit trail of what it did.
Human oversight
Production agents we build include a review step before any action with financial, legal, or customer-facing consequence, sized to the risk of that action-full sign-off for high-stakes actions, sampling for low-stakes ones.
Data handling in model workflows
Where a client's data is used to fine-tune or ground a model, it stays inside that client's own environment or an isolated tenancy-never pooled with another client's data, and never used to train a shared or public model without explicit written agreement.
Evaluation before production
Every model-backed system we deliver is evaluated against defined accuracy, safety, and failure-mode benchmarks before it reaches production, and re-evaluated when the underlying model changes.
Reporting a concern
If you believe a system we built is behaving unsafely, tell us-email ai-safety@shellkode.com with what you observed. We treat these reports with the same urgency as a security report.
Contact us
ai-safety@shellkode.com for anything specific to how our AI systems behave; security@shellkode.com for vulnerabilities in the systems themselves.