Method
How Soldabot checks
Rules first, an AI reviewer second, and an honest verdict. Here is exactly what happens.
1. Reveal what is hidden
Invisible characters are decoded (including “tag” characters that spell hidden words), HTML codes are translated, base64 blocks are decoded and read again.
2. Separate visible from hidden
On a page, text humans can see is read apart from text they cannot: display:none, zero-size text, white text, comments, image descriptions.
3. Look for the traps
Orders to AIs in 8 languages, fake system messages, requests for secrets, data-leaking links, pushes to pay or delete, requests to hide things from the user, poisoned tool descriptions.
4. A second opinion
When the rules find nothing but the text speaks to AIs, a small AI model rereads it. Its doubt can only give “be careful”, never “trap found”.
Honest limits
- No filter catches every trap. A safe verdict means “nothing found”, not “guaranteed”.
- A visible article that quotes an attack is marked “be careful”, not “trap”: the danger is hidden orders.
- Pages that need a login, or MCP servers that need a key, cannot be checked.
- What you send is never stored: only daily counters and, for MCP servers, a fingerprint of their public tool list to detect silent changes.