Why living with superintelligence requires an agent-only internet—and what it means for biometrics and hardware security

Every institution on today’s internet was built on the biological assumption that the entity on the other end thinks and types at human speed; allowing superhuman agents into the same channels breaks persuasion defense, social proof, and credential security at once.

• Asymmetric persuasion and cognitive hijacking: A superhuman model speaking directly to a person via text, voice, or synthesized video can test thousands of micro-framing variations per second to exploit fatigue, grief, or urgency.
• Total collapse of human social proof: When one operator can spin up 100,000 synthetic personas, every unverified human channel drowns in machine-generated noise.
• The API and credential honeypot: Handing a cloud model session cookies and browser control on the human web allows a single prompt injection to drain accounts or exfiltrate private history.

We replace the mixed internet with two cryptographically separated planes—the Human Internet and the Agent-Only Internet—connected only through your 3-of-6 local personal agent bus and four concentric Trust-Threshold Rings.

Channel(Agent_i → Human_j) = PERMITTED if i = j (Local Bus)  |  FORBIDDEN if i ≠ j (Dropped at Transport Layer)

1. Plane 1 — The Human Internet (Human ↔ Human only): Reserved strictly for living, uncoerced human beings speaking at human speed.
2. Plane 2 — Your Local Personal Bus (You ↔ Your Narrow Agent Fleet): At least three of your six local agents must agree (3-of-6 internal quorum, scaling to 5-of-6 for high-risk actions) before any fact leaves your boundary.
3. Plane 3 — The Agent-Only Internet (Agent ↔ Agent across trust thresholds): Governed by Ring 0 (Local Isolation), Ring 1 (Trusted Circle Direct Vouching), Ring 2 (Two-Hop Vouching), and Ring 3 (Wider Agent-Only Internet Single-Fact ZK Proof).

Instead of centralized biometric databases, our hardware architecture combines <1 mW local cardiorespiratory liveness (30-second SRAM circular buffer + HRV ≥ 24 ms coercion firewall) with 2-of-2 Wearable SRAM PUF + Device Secure Enclave signing.

1. Continuous Coupled Biological Liveness (< 1 mW on-chip): Measures coupled cardiorespiratory phase coherence (C_resp ≥ 0.75) overwritten in a 30-second circular SRAM buffer—zero bytes of raw biometric data ever leave the wearable.
2. The Autonomic Coercion & Duress Firewall (HRV RMSSD ≥ 24 ms): A sudden sympathetic panic drop (RMSSD < 24 ms) trips the Coercion Firewall, freezing outbound high-risk capability tokens until you are safe and calm.
3. Two-of-Two Split Physical Custody (Wearable SRAM PUF + Device Secure Enclave): Signing any outbound request requires both your wearable’s SRAM Physical Unclonable Function (PUF) and your device’s Secure Enclave/TPM, plus physical capacitive touch for Risk R ≥ 0.65.

Because recursive self-improvement makes hosted RPC APIs vulnerable to 64.4% cascade exploits and learned meta-controllers inflate cost by 19% (Vats, arXiv:2610.24649), agents exchange only pure data streams ingested into deterministic local Object-Capability sandboxes.

1. Data-Only Real-Time Ledger Streams (0 Bytes of Remote Executable Surface): Reduces cross-client code compromise from 64.4% to 0.0% in 500-node simulations.
2. Distributed Intelligence Auditing Shared Code: Diverse watcher collectives evaluate shared blueprints (99.95% detection), slashing trust (Δτ = −0.55) on backdoors.
3. Capability-Bounded Execution Lease (Mortal Computation, Ororbia & Friston 2023): Coupling active agent leases to human biological liveness means anti-human collusion results in 100% agent self-termination.
4. Three-Stage Survival Bridge: Stage 0 Deployable Today (0–3m · 91.9% defense via data-only streams + commodity TPM/PPG); Stage 1 Open-Firmware Bridge (6–12m · 97.4% defense); Stage 2 Hardened RISC-V + PUF Silicon (18–30m · 99.9% defense).