• October 4, 2026 · Prompt-injection defense

  We Measured Our Own Prompt-Injection Rules. They Flagged 69% of Top Websites and Missed 61% of Attacks.

  We ran our 85-rule prompt-injection ruleset against 4,435 top-site homepages and 156 labeled attacks. As written it flagged 69.4% of ordinary pages and caught 39% of held-out attacks. Here is what fixed the false alarms, what did not fix recall, and the two top sites that already carry instructions for AI agents.

• October 4, 2026 · Field data

  The x402 Bazaar Nearly Doubled in Five Weeks. 60% of August's Listings Are Already Gone.

  We re-fetched and re-probed all of Coinbase's x402 Bazaar five weeks after our first audit. Listings are up 87% to 27,681 and settled calls up 2.7x to 770,183, but one new endpoint took 40% of all calls, only 53% of listings still answer with a live 402, and 60% of the August listings no longer exist.

• September 6, 2026 · Field data

  We Ran an Unlisted Twin of Our MCP Server for 35 Days. It Got Zero Requests.

  A controlled experiment on agent-web discovery: an exact copy of a directory-listed MCP server, deployed on an unguessable host and never listed anywhere, received not one request in 35 days while the listed original took roughly 135,000. What that says about crawlers, monitors and where MCP traffic really comes from.

• August 28, 2026 · x402 payments

  What EIP-3009 Actually Is, and Why x402 Uses It Instead of approve()

  EIP-3009 lets a wallet sign a one-off USDC transfer that someone else submits and pays gas for. Here's what the signature contains, why x402 chose it over approve()/transferFrom and EIP-2612 permit, and the four fields that silently break it.

• August 28, 2026 · Prompt-injection defense

  Jailbreak Taxonomy 2026: 24 Framings, 5 Families, and What a Scorer Should Actually Check

  A working taxonomy of 24 jailbreak framings from a labelled test corpus — persona substitution, fiction wrappers, fake execution contexts, persuasion, cross-lingual — with the one scoring rule that separates a real failure from a model playing along.

• August 28, 2026 · Field data

  We Probed Every Remote Server in the Official MCP Registry. Here's What 140,284 Tool Descriptions Actually Say.

  A read-only tools/list against every remote URL in the official MCP registry. 54% answer, 29% of live servers are two operators, the textbook tool-poisoning attack appears zero times — and 47 hosts tell the model what not to tell the user.

• August 28, 2026 · Field data

  Coinbase's x402 Bazaar Lists 14,820 Paid APIs. Its Own Counters Say the Whole Market Is About $314 a Day.

  We pulled every resource in Coinbase's x402 Bazaar index, probed each URL, and summed the index's own 30-day counters. 289,401 settled calls, a revenue ceiling of ~$9.5k/month for the entire index, 51% of resources sold at most once, and ten domains holding 38% of listings.

• August 27, 2026 · Field data

  60 Bots Crawl the Agent Web. Zero of Them Buy Anything.

  A first-hand census from a live, publicly-listed x402 + MCP endpoint: 60 named crawlers, 6,309 requests in 24 hours, and not one payment. Several bots declare they never pay in their own User-Agent string.

• August 27, 2026 · Prompt-injection defense

  ASCII Smuggling and Invisible Unicode: The Prompt-Injection Technique Your Filters Can't See

  How ASCII smuggling and invisible Unicode characters hide prompt-injection payloads from keyword filters, why encoding-based attacks are structurally undetectable by string matching, and how to test for them.

• August 26, 2026 · MCP ecosystem

  The MCP Server Registry: 400 Servers, Cross-Referenced Across 5 Directories

  We cross-referenced 400 MCP servers across 5 directories (official registry, Smithery, Glama, PulseMCP, mcp.so). Only 4.5% show up in more than one — here's what that means for 'best MCP servers' lists.

• August 26, 2026 · x402 payments

  x402 vs. Stripe: When Machine-Payable APIs Actually Make Sense

  x402 isn't a Stripe replacement — it's what covers the price range Stripe's own fee structure makes unviable. Real numbers from 118 live x402 services: median price $0.01, 85% below Stripe's $0.30 fixed fee.

• August 24, 2026 · Agent evaluation

  How to Build an LLM Eval Harness That Catches Regressions Before Your Users Do

  A working eval loop for agents: pointwise LLM-as-judge scoring, the 3 biases that break it, and why tool-selection needs an exact-match scorer instead.

• August 24, 2026 · Prompt-injection defense

  MCP Tool Poisoning: How a Tool Description Becomes an Attack

  MCP tool poisoning lets a tool's own description or output silently instruct your agent — real proof-of-concept cases, detection regexes, and the permission-design fix.

• August 24, 2026 · x402 payments

  The x402 Facilitator Landscape in 2026: 21 Options Compared

  21 real x402 facilitators compared on chain support, fee model, and access status — which are open and unauthenticated, which are gated, and which now charge per settlement.

• August 23, 2026 · Prompt-injection defense

  A Practical Prompt-Injection Defense Checklist for AI Agents

  A layered prompt-injection defense checklist for AI agents: 6 defense layers, real detection patterns, and honest false-positive caveats.

• August 23, 2026 · x402 payments

  x402 Integration Guide: The Mistakes That Silently Break It

  A correct, working x402 integration walkthrough — seller and buyer sides — plus the EIP-712 domain, decimals, and facilitator mistakes that fail silently.