Archive
13 artifacts published in this period.
The same 250-document backdoor that compromised LLMs from 600M to 13B params, plotted across model scale. Toggle 'documents needed' — a flat 250 vs the myth that poison scales with data — and 'share of training set', collapsing to 0.00016% and below. Drag to pick a scale.
How many audit probes it takes to catch a substituted model, as a function of how subtle the swap is. A 1/Δ² curve over real quantization accuracy gaps: a model swap is caught in a few hundred probes; FP8 needs ~150k and falls off the cliff into the economically-invisible zone.
Watch a searcher sandwich your agent's swap on the live Aerodrome WETH/USDC pool on Base. Set trade size and slippage tolerance; the optimal front-run is solved on the real constant-product curve, capped by your tolerance. Route privately and the extraction goes to zero.
Allora weights models by regret, not votes. Drag a market from calm to a volatility spike: five forecasters ride the softplus weight curve φ′(R̂), the low-vol specialist collapses the instant a spike is forecast while the vol-aware model climbs. Tap a model for its R→R̂→w chain.
Does an AMM LP out-earn the arbitrage bots picking off its stale quotes? Plot fee income (flat) against LVR = σ²/8 (rising with vol²): where they cross is break-even. At low turnover even BTC sits in the red. Flip on solver-auction recapture and the frontier slides right.
Open weights can't be locked, so ownership is proven by counting surviving fingerprints. Embed thousands of secret key-response pairs; an adversary fine-tunes to scrub them and each audit burns one to leakage. Watch the reserve survive — or not.
A single mech request, drivable: drag the priority mech's response time across the 60-300s window. Inside it the priority mech delivers and gains Karma; past it a backup delivers, the priority mech forfeits the fee and is docked Karma. Live Olas marketplace constants.
Drive a Virtuals ACP escrow job: client funds USDC, provider submits a deliverable hash, and whoever sits in the evaluator's chair springs the payout. Seat the client (what ~all of Base does), a neutral agent, or no one — then run it or let someone cheat, and watch who's left exposed.
Pick the GPU a provider bills for and the cheaper card it secretly runs; the network measures the real device's compute, bandwidth and VRAM and plots it against the claimed model's fingerprint. Watch the spoof land outside the tolerance band — or hide inside it.
A liquidation mints a fixed prize — the bonus on seized collateral. Three bars split it three ways: a gas war bleeds it to the block builder, private orderflow pockets it for the searcher, an OEV auction returns ~73% to the protocol. Drive size, bonus, competition; anchored to Chainlink SVR figures.
An AI proposer is wrong a few percent of the time, but UMA pays a winning disputer only half the bond it risks — so the break-even belief is two-thirds, not half. Every category's base error rate sits deep in the no-dispute zone; raise the reward and the threshold slides left.
When an LLM agent writes the exploit, the fight is unit economics. Break-even contract value against cost-per-scan: the attacker's line, and the defender's sitting 10× higher because a bounty pays a tenth of a theft. The band between is the attacker-only zone, with A1's six models at measured cost.