Audits, exploits, threat models, and defensive engineering across the stack.
You pay per token for a named model; the provider picks the precision. FP8 quantization costs 0.6 MMLU points and is near-invisible to output auditing — so inference markets bond and attest instead of detect.
A landmark 2025 result: ~250 poisoned documents backdoor an LLM whether it has 600M or 13B parameters — 0.00016% of the tokens. DataDAOs sell 'verifiable' training data, but on-chain provenance proves integrity, not purity. Here's the gap, and what actually narrows it.
Permissionless training networks pay peers for gradients they can't re-run. Proof-of-Learning was meant to verify the work — until it was forged for 3% of the training cost, then for one floating-point op per weight. Here's the mechanism, the attacks, and what's deployed instead.
Wire an LLM into a Governor contract and the proposal text becomes the attack surface. A frozen-weights ablation finds the twist: turning on deliberation drops adversarial robustness from 100% to 68.5% — and the 226s it takes to think is itself extractable value.
An agent doesn't need its key stolen to drain a wallet — it can be talked into signing. CrAIBench shows memory injection beating prompt injection 55% to ~0% on the strongest model, and only fine-tuning closes the gap.
zkML, optimistic, TEE, FHE all prove the computation. Restaking takes the other road: bond it and slash liars. We do the cost-of-corruption math behind EigenLayer's $18B AI-AVS security, the overloading attack that breaks it, and the probabilistic-audit tax.
An agent holding your raw key is one prompt injection from total loss — and 97% of early EIP-7702 delegations went to drainer sweepers. We read the sweeper's source off the chain, dissect a real 35.97-USDC-a-day spend permission on Base, and do the blast-radius math.
World Chain reserved top-of-block space for verified humans while 382M smart-account ops poured in. We read the Semaphore tree on Ethereum — 17.59M identities, hourly batches at ~3,300 gas each — decode PBH's month-stamped nullifiers, and find the human-only lane sitting almost empty.
The Elliptic benchmark made GNNs the default for on-chain AML. A 2026 leakage-free re-evaluation flips the script: random forests win by 13 F1 points, randomly rewired edges beat the real graph, and every model falls off a cliff at time step 43.
ERC-8004's agent registries went live in January. We read them straight off the chain: ~90,000 registrations across Ethereum and Base, a $0.003 ERC-4337 registration dissected — and a reputation registry already farmed with vouch spam.
zkML costs 1000x, optimistic schemes cost a challenge window. Hardware attestation verifies AI inference at under 7% overhead and ~$0.26 on-chain — if you're willing to trust Intel. Part 3 weighs the third leg of verifiable inference.
An autonomous agent can't sign up for an API key or swipe a credit card. x402 revives HTTP 402 and settles in gasless USDC — here's the EIP-3009 handshake, the facilitator trust model, and where prompt injection breaks it.