TIBET Cortex

Zero-trust AI knowledge processing. Data that protects itself.

After CVE-2026-0866 (Zombie ZIP) showed that 50 out of 51 AV engines can be fooled by header manipulation, and McKinsey's Lilli breach exposed 46.5 million chat messages through a single SQL injection, one thing is clear: the data itself must be the security boundary, not the application around it.

TIBET Cortex is a Rust framework for building AI knowledge systems where every document chunk protects itself — cryptographically, at every layer, in every state.

Architecture

┌─────────────────────────────────────────────┐
│              TIBET Cortex                    │
│                                              │
│  STORE     TBZ envelopes + JIS levels        │
│            Embedding JIS 0 (searchable)      │
│            Content JIS N (encrypted)         │
│                                              │
│  GATE      Multi-dimensional JIS claims      │
│            role × department × time × geo    │
│                                              │
│  AIRLOCK   Zero plaintext lifetime           │
│            mlock + zeroize + scope-bound     │
│                                              │
│  AUDIT     Blackbox-met-window               │
│            See WHO/WHEN/HOW MUCH, not WHAT   │
│            Immutable TIBET provenance chain   │
│                                              │
│  VAULT     Time-locked audit trails          │
│            Dead man's switch for compliance   │
└─────────────────────────────────────────────┘

The Problem

Every RAG stack today:

User Query → Embedding → Vector Search → ALL docs → LLM → Response

No layer between "found" and "allowed to read." One SQL injection, one IDOR, one misconfigured endpoint — and your entire knowledge base is exposed.

The Solution

TIBET Cortex separates search from access:

Embeddings are JIS level 0 — always searchable
Content is JIS level N — cryptographically gated
Processing happens inside an Airlock — zero plaintext lifetime
Every operation generates a TIBET audit token

SQL injection dumps the DB?
→ You get encrypted TBZ envelopes. Unreadable.

Memory dump during processing?
→ Airlock wiped. mlock'd memory zeroed.

Audit trail tampered?
→ TIBET chain broken. Immediately detectable.

System prompt modified?
→ Integrity hash fails. Execution refused.

Crates

Crate	Description
`cortex-core`	TBZ envelopes, TIBET tokens, Ed25519 crypto, zeroizing buffers
`cortex-airlock`	mlock'd memory, scope-bound processing, auto-wipe
`cortex-jis`	Multi-dimensional claims: clearance × role × dept × time × geo
`cortex-store`	sled-backed JIS-gated vector storage
`cortex-audit`	Blackbox-met-window audit trails, TIBET chain verification
`cortex-cli`	Command-line interface

Install

cargo install cortex-cli

Usage

# Ingest a document at JIS level 2 (confidential)
cortex ingest ./strategy.pdf --jis-level 2 --source strategy-db

# Query with your identity claim
cortex query "M&A targets" --clearance 3 --role partner --department strategy

# Verify audit chain integrity
cortex verify

# View audit statistics (blackbox-met-window)
cortex audit
cortex audit --full

# Show architecture
cortex info

JIS — Multi-Dimensional Access Control

JIS is not a single number. It's a multi-dimensional identity claim:

let claim = JisClaim::new("partner@mckinsey.com", 3)
    .with_role("partner")
    .with_department("strategy")
    .with_geo(vec!["NL".into(), "DE".into()]);

let policy = JisPolicy::clearance(3)
    .with_roles(vec!["partner".into()])
    .with_departments(vec!["strategy".into()])
    .with_geos(vec!["NL".into(), "DE".into(), "FR".into()]);

// All dimensions must match
assert!(JisGate::is_allowed(&claim, &policy));

An intern in the US sees different data than a partner in the EU — from the same query, on the same system.

Airlock — Zero Plaintext Lifetime

let airlock = Airlock::with_defaults();

let (result, audit) = airlock.process(
    encrypted_content,
    "analyst@company.com",
    2,  // JIS level
    |plaintext| {
        // This closure is the ONLY place plaintext exists
        // Memory is mlock'd (never swapped to disk)
        Ok(process(plaintext))
    },
)?;
// After closure: all plaintext memory zeroized
// audit token generated for the trail

Blackbox-met-Window Audit

The auditor sees:

WHO accessed data
WHEN it was accessed
HOW MUCH data was touched (chunks accessed/denied)
WHAT JIS level was used

The auditor does NOT see:

The actual content
The query itself (only its hash)

Unless they have matching JIS credentials to resolve the hashes.

Benchmarks

Run it yourself:

git clone https://github.com/jaspertvdm/tibet-cortex.git
cd tibet-cortex
cargo test --release --test bench_airlock -- --nocapture

What's measured

Test	What it measures	Iterations
`bench_airlock_single_chunk`	Airlock overhead for one 4KB chunk: `mlock` → `memcpy` → closure (SHA-256 hash) → `zeroize` → `munlock`	1,000
`bench_airlock_batch_chunks`	Same pipeline for 10 chunks (typical RAG response), processed sequentially in one Airlock session	100
`bench_full_search_pipeline`	End-to-end: scan 100 embeddings + JIS gate evaluation + cosine similarity ranking + Airlock processing of top-5 results	100

Timing uses std::time::Instant (monotonic clock). No warmup runs — first iteration included. Assertions are relaxed for debug mode; the numbers below are release mode.

Reference numbers

Metric	Latency	Throughput
Airlock single chunk (4KB)	134µs	7,436 ops/sec
Airlock batch (10 × 4KB)	611µs (61µs/chunk)	16,374 chunks/sec
Full search (100 docs, top-5)	1.5ms	665 searches/sec

Measured on: Xeon E5-2650 v3 @ 2.30GHz, 64GB DDR4, Linux 6.12. Your numbers will vary — mlock latency depends on kernel config, available RAM, and RLIMIT_MEMLOCK.

Context

A typical LLM response takes 500ms–2s. The entire zero-trust pipeline (envelope unpacking, JIS gate, cosine search, Airlock) adds ~1.5ms — less than 0.3% overhead.

Part of the TIBET Ecosystem

TBZ — Block-level authenticated compression
TIBET — Trust & Identity Blockchain for Ethical Transactions
OomLlama — Rust-native LLM inference engine

License

MIT OR Apache-2.0

tibet-cortex-core 0.3.0