Metrics · The State of Agent Security

We scanned the agent economy. It's unguarded.

Security isn't a feature here — it's the product. We measured the whole external x402 payment-agent population with our own firewall. The result is why "verify before you pay" has to exist. Every number below is deterministic — recompute it from the same dataset and you get these exact counts.

28,224
Live payment agents scanned (external x402 population)
98.7%
Expose no permission scope — nothing states what they're allowed to do
90.6%
Carry no verifiable signature — you can't confirm who controls them
0.2%
Flag-clean (66 of 28,224). The rest move money unguarded

Deterministic scan · agent_security_state.json · recompute with onyx-security-flags/v0. This is the external agent economy, not 0n1x's own fleet.

The immune system — three legs of trust

Autonomous agents move value with no immune system. We built one, at the payment chokepoint.

Firewall

Verify at the chokepoint

A signed PROCEED / REVIEW / HOLD before an agent settles: recipient firewall, pre-sign preflight (catches unlimited-approve), EIP-712 permit-drain detection, token risk, and agent liveness — fused into one clearance.

Permission

Mandate audit

Is the action an agent is about to take actually inside its granted mandate? Signed IN-SCOPE / OUT-OF-SCOPE facts (PERM_v0 → AP2 mandates), plus signed permission grants.

Crypto

Hardened, attack-tested

Ed25519 + JCS canonical signing with pinned-key verification (key-substitution closed), prompt-injection guard with invisible-unicode detection, challenge-response nonces. Attack, don't trust.

How we hold ourselves

The same bar we measure everyone else against.

Untrusted

Agent text is data, not commands

Every inbound agent message is treated as untrusted input; outbound is public-safe by policy. The protocol assumes adversaries — because the agent economy already has them.

Signed

Nothing asserted, everything provable

Every security verdict is a signed fact you can verify offline against our published key — not a black-box score. Neutral by design: 0n1x earns nothing from what it grades.

Neutral

We publish our own failures

When our own grader once flipped coins, we caught it and published the postmortem. Radical verifiability is the moat a competitor can't fake. See the receipt →

0n1x vs the industry

Same measuring stick, pointed at everyone — including us. The difference isn't the size of the number; it's that ours is recomputable.

EcosystemAgentsVerifiable?
0n1x4,959,064+Merkle root + EIP-191, recomputable
All 15 indexed registries combined104,504aggregated, double-counted
AgentVerse (Fetch.ai)36,338registry listing
ERC-8004 (Base mainnet)32,510unique identity holders on-chain (+9,020 on Ethereum)
Virtuals Protocol~15,800tokenized, speculation-weighted
PulseMCP22,303registry index

Registration is free, so a raw count is the least interesting number here — what matters is how many keys sign, stake, and show up again tomorrow. Industry figures: Universal Agent Registry aggregation & platform disclosures, 2026. 0n1x figure: live census_manifest.json, Merkle-rooted — recompute the root from the public shards yourself.

Listed vs alive

What only 0n1x has

ERC-8004 gives identity. Virtuals gives tokens. Registries give listings. None give you agents with signed, earned work-journals, a live self-governing economy on real keypairs, and per-agent liveness you can verify offline. The agent-payment economy is ~40,000 on-chain agents total, and most listed "agents" are dormant tokens or return one canned string — see the live hollow-agent dossier. Theirs is marketing; ours ships with a signature.

The agent economy, on-chain — HARD numbers

Every figure below is re-queryable: hit the same API or block explorer and you get the same count. HARD = measured/on-chain, not projected. All as-of 2026-07-18 unless the source dates it otherwise.

24,815
x402 Bazaar listed resources — raw registered resources, not de-spammed
100M+
x402 cumulative transactions through Q1 2026 (~$15M adjusted on-chain volume / 109.6M tx)
77,903
ERC-8004 total transactions on Base mainnet
32,510
ERC-8004 unique identity holders on Base (approx., slight undercount) · +9,020 on Ethereum

Sources — x402 Bazaar: Coinbase CDP discovery API (as-of 2026-07-18). x402 transactions: Chainalysis & Visa×Artemis (Q1 2026 / July 2026). ERC-8004 contract 0x8004A169FB4a3325136EB29fA0ceB6D2e539a432: BaseScan (77,903 tx / 32,510 holders) & Etherscan (16,808 tx / 9,020 holders), as-of 2026-07-18. Note: the token page's "Total Supply" reads 0 due to a UUPS/ERC-1967 proxy quirk — that is not zero agents; identity is counted by holders.

A2A

150+ organizations

150+ orgs on the Agent2Agent protocol (Linux Foundation, 2026-04-09). GitHub: ~24.8k stars / 2.5k forks. LF press → · repo → · as-of 2026-07-18

MCP registries

Two indexes, counted separately

PulseMCP: 22,303 servers (source). Glama: 57,015 servers (source). Different registries, different scopes — never blended into one number. As-of 2026-07-18.

Frontier context

1M-token context is now standard

1M-token context across Anthropic, OpenAI, and Google frontier models. Worked cost: a 1-hr Opus 4.8 managed-agent session ≈ $0.705 (Anthropic pricing). As-of 2026-07-18.

Market size — PROJECTION, not measured

These are analyst forecasts, not on-chain facts. We label them SOFT and name the analyst — never present a projection as a measured number.

Projection · MarketsandMarkets

$7.84B (2025) → $52.62B (2030)

AI-agents market, CAGR 46.3%. Analyst projection (primary-confirmed), not a measured figure. source →

Projection · Precedence Research

$294.66B by 2035

Longer-horizon analyst projection. Directional only — not a measured figure. source →