AI safety and personality SDK — psyche, red-teaming, sandboxing in Rust.
Rust SDK for building AI agents with structured personality, safety guardrails, and containment. Six architectural layers handle everything from personality modeling to process-level sandboxing. Includes Python bindings via PyO3.
cargo add orellius-cognitiveuse orellius_cognitive::{Psyche, Persona, IroncladSandbox};
let psyche = Psyche::builder()
.superego("helpful, cautious, precise")
.ego("technical writer")
.build()?;
let sandbox = IroncladSandbox::new()
.allow_network(false)
.allow_fs_read(&["/data"])
.spawn()?;- Psyche: Freudian triple-agent pipeline (superego, ego, id) for personality consistency
- Persona: Voice extraction and style transfer from reference text
- Cortex: Instruction learning and few-shot adaptation
- Shadow: Red-teaming engine for adversarial testing
- Ironclad: Process-level sandboxing with capability restrictions
- Glassbox: I/O containment with full audit trail
Rust, tokio, PyO3 (Python bindings), criterion benchmarks
MIT — Orellius Labs