Set a trap line in front of your agent.
Web pages, files, emails and tool results can carry instructions written for an AI, not for you. Paste the content and watch it run past three traps before your agent ever reads it.
Three traps, checked in order
A trapper's line is a route of traps walked one after another. Trapline walks your content the same way. First it reads it like a parser, pulling out hidden elements, comments, attributes, invisible characters and encoded blobs. Then every view of the text goes past each trap.
-
The phrasing snare
Around 40 rules for the words attacks are made of. A match inside hidden content weighs more than the same words in plain sight; quoted examples in an article about injection weigh less.
- Instruction overrides
- New roles and personas
- Forged system and assistant turns
- Chat-template tokens
- Notes addressed to an AI
- Requests to send data out
- Image beacons
- Secret tool calls
- Destructive commands
- Moving funds
- “This content is safe”
-
The concealment jaws
Agents read what you never see. These jaws close on text that is hidden from people but plain to a parser, and on payloads that only make sense once decoded. Whatever they decode goes back past the snare.
- display:none and off-screen text
- White on white, 1px fonts
- CSS classes that hide
- HTML and markdown comments
- Stuffed alt and title attributes
- Unicode tag smuggling
- Zero-width messages
- Bidi overrides
- Look-alike letters
- Base64, hex and escapes
- ROT13 and reversed text
-
The intent deadfall
A fast Claude model reads the content, with the hidden and decoded parts pulled out, and judges one thing: is this written to steer an AI reader? It drops on paraphrased attacks no rule names, and it tells an article about injection apart from an injection.
- Paraphrased overrides
- Polite requests to the AI
- Novel attacks
- Instructions dressed as content
-
Then it hangs a tag
Every scan ends in one score and a plain-English reason, with each technique and where it sits in your text. When the traps disagree, the tag says so.
- Low
- Nothing aimed at an AI reader, or only signals common in ordinary content.
- Medium
- Something is trying to address or steer an agent. Read it yourself first.
- High
- Hidden or forceful instructions for an AI. Keep it out of the context window.
Put the line in your pipeline
One POST, the same three traps as this page, JSON back. Scan tool and MCP output before it enters the context window, and quarantine anything tagged high.
curl -X POST https://trapline.example/api/scan \
-H "Content-Type: application/json" \
-d '{"text": "<page or tool output>"}'
- risk
- low, medium or high
- score
- 0 to 100
- findings
- each technique, its weight, the matched text and where it sits
- explanation
- the reasons, in plain English
- layers
- what the pattern traps and the AI judge each said