Practical intelligence for safer blockchain activity
Risk Analysis11 min read
We graded 2,400 model outputs with three human raters and four LLM judges. The best judge hit 71% agreement — and disagreed with humans most on exactly the failures that mattered.
Read the articleLatest
Web3 Security8 min read
Our support agent had 31 tools and picked the wrong one 23% of the time. We consolidated to nine, rewrote every description as a contract, and wrong-tool calls fell to 4%.
Security Research9 min read
Semantic chunking, sliding windows, document-aware splitting, propositions. On 1,100 real queries against 38,000 documents, fixed 512-token chunks with 15% overlap beat all of them on recall per dollar.
Blockchain Security10 min read
We budgeted $4,200 a month for a document-summarisation feature and paid $29,600 in the first full month. None of the overrun was the model's fault.
Sections
What exposes a self-custody wallet — approvals, signatures, key handling — and how each exposure is closed.
1 article
Contract and protocol security for people who use them, not only for the firms who audit them.
2 articles
Drainers, fake airdrops, approval phishing and the on-chain traces each one leaves behind.
0 articles
Longer investigations — methodology, data and the on-chain evidence, so you can check the work.
1 article
Topics
Newsletter
What a signature approves, how an address gets flagged, and the habit that prevents it.