AIKIDO-2026-631787

TL;DR

The local web_fetch tool converts fetched HTML to Markdown by processing nested block elements and reprocessing each descendant's accumulated text at every level, so a small page with deep nesting such as repeated blockquotes, list items, or definition items expands into a very large intermediate string. Fetching untrusted HTML with this structure consumes excessive CPU and memory and stalls other work in the process, and on older releases the conversion ran directly on the event loop, leading to denial of service. The response body limit only bounds downloaded bytes and the returned content limit applies after conversion, so neither bounds the cost of the conversion itself. The fix adds conversion cost, text scan, and nesting depth limits that abort the conversion once the work budgets are exceeded.