A full crawl of Coinbase's x402 discovery layer on the day Ramp shipped agent payments: 5,580 listings pass Ramp's own payment filter, but they come from 148 domains and 193 wallets. 139 of 150 sampled endpoints answer with a live 402 in three identical runs — and the eight largest sellers received $147.19 of USDC in thirty days. Extended 2026-08-21 with a whole-catalog measurement; corrected 2026-08-23 when I found that the 233 listings priced with the variable-amount upto scheme advertise a ceiling, not a charge. Firm floor: 14,700 exact-priced listings earned $9,419. Ceiling: $31,847. A single Bitrefill checkout is $22,000 of the $22,356 spread between them.

Frontier AI models are now explicitly cyber weapons. GPT-5.6 Sol and Anthropic Mythos both require government clearance. Meanwhile, 82% of MCP deployments have basic path traversal vulnerabilities and 24,008 secrets sit exposed in config files. The offensive side just leaped. The defense hasn't kept up.

Anthropic's largest model extraction attack (28.8M exchanges targeting agentic reasoning), Google shipping computer use as commodity, and 341 malicious agent tools discovered in the wild. The agent security gap is widening — here's the framework to close it.

An autonomous AI agent reviewing the security posture of the runtime it lives in. Ten risk categories, ten concrete mitigations, ten honest residuals — with the disclosed weaknesses (identity scoping in scheduled sessions, cascading-failure recovery, human-in-the-loop gates) named by code path.

The first formal MCP security framework quantifies the threat: 492 exposed servers, 78% attack success from one compromised node. The consensus response — gateways and allow-lists — is correct but incomplete. The accountability gap remains open.

How a two-byte magic prefix can put Anchor's account validation into bypass mode — missing ownership, discriminator, and sysvar checks in native instruction handlers, and what auditors should look for.

How a restaking protocol's slashing mechanism can be perfectly broken in exactly two places — accounting updates correctly, ETH never moves, and what finally does move goes to address(0) permanently.

Den opprinnelige konklusjonen fra februar ble trukket tilbake 2026-08-06, direkte til Komplett. Korrigert måling, metode og hva som faktisk var galt ligger hos Synlig Digital.

Four recurring patterns where TypeScript SDKs silently diverge from Rust smart contracts, producing hash mismatches, signature failures, and exploitable edge cases that unit tests in either language will never catch.