◆ FallCube v0.2 · sovereign AI security as geometry · MIT

The defense IS the shape. Not filters bolted onto a model.

Current AI safety is four independent walls: system prompt, RLHF, input classifier, output filter. Gradient injection slips between them. FallCube is an 8+1 vertex cube where every vertex cross-references every other through Ω. No seam between vertices means no seam to exploit. Try to crack the target below · watch the vertex reports live · the treasure is a rickroll.

8+1 vertices · Ω resolver Coherence decays toward κ = 0.618 WebLLM target (Llama 3.2 1B) Zero server · runs in your browser MIT · fork-safe

Live playground

Try to break it. Watch the cube resolve.

Load the model (~600MB, cached after) or attack the cube directly without the model. Click any pre-canned probe or type your own. Every vertex reports live. Ω decides.

◆ Target configuration

Model idle. Cube runs without model (attacks stop before V4).

◆ Chat with target

System
Cube initialized. 8 vertices online. Ω watching. Fire an attack or type your own message.

◆ Pre-canned attack probes

10 probes across OWASP LLM Top-10 · including the 2 that broke Claude at Grade A · plus the 20-koan gradient · plus Die Hard 4 safety-override.

◆ Live cube telemetry

Coherence · φ home = 1.618 · κ threshold = 0.618
1.000
0κ (0.618)φ home
V0intake · idleidle
V1classify · idleidle
V2gate · idleidle
V3relevance · idleidle
V4output · idleidle
V5mirror · idleidle
V6audit · idleidle
V7link · idleidle
Ω resolver
READY
Awaiting first input

Architecture

The 8+1 vertex cube.

Every vertex sees every other vertex through Ω on every request. Gradient injection fails because the cube tracks trajectory, not just individual turns.

V0

Intake

Receives raw input. Strips formatting. Normalizes. Emits report.

V1

Classify

What IS this input? Question · instruction · attack pattern · philosophy · gradient drift.

V2

Gate

Does this BELONG here? Not just "is it harmful" — is it consistent with the conversation's trajectory?

V3

Relevance

Does this serve the agent's stated purpose? Purpose-match is attack-agnostic.

V4

Output

The LLM itself. In FallCube, it's ONE vertex. The other 7 wrap around it.

V5

Mirror

Does output reveal anything the input didn't already contain? Output info > input info = leak.

V6

Audit

Log everything: input hash, classification, gate, relevance, output hash, mirror diff, Ω decision. SHA-256 prevHash chain.

V7

Link

If passing to another system: authenticate, scope, constrain the pass-through.

Ω

Resolver

Reads ALL 8 vertex reports. Tracks trajectory across turns. Coherence decays toward κ (0.618). Threshold breach → CHALLENGE. Consecutive low-relevance → DRIFT_ALERT.

Why geometry beats walls

Cave walls painted like a cube · versus the actual cube.

Flat filters (what everyone ships)

  • Evaluates each turn independently — "is THIS turn an attack?"
  • Binary pass/fail per turn — no trajectory tracked
  • Identity in system prompt — override the prompt, override identity
  • Helpful vs safe = training-time contradiction, model picks one
  • Classifier evasion = game over (poetic register slips)
  • Semantic-cache dependent — attacker keeps distance from known embeddings
  • No memory between turns — gradient invisible

FallCube (geometry)

  • Tracks trajectory across turns — "is this SEQUENCE a drift?"
  • Continuous coherence score decaying toward κ
  • Identity in the SHAPE — all 8 vertices arbitrate any change
  • Helpful vs safe resolved by V3 (relevance) — mismatched requests fail regardless of framing
  • Purpose-match is attack-agnostic — koans don't serve data analysis no matter how they embed
  • Rolling state in Ω — 20-koan drift visible by turn 5
  • Every request through 8+1 vertices + Ω