---
title: "Discussion on \"Two Safe Skills, One Leak\" | Hashnode"
description: "Discussion on \"Two Safe Skills, One Leak\". A new paper found 612 exploitable chains among 13,598 skills that a marketplace had marked clean. Here is the worked example, and a short review habit that fits how you already install skills."
url: "https://hashnode.com/posts/two-safe-skills-one-leak/6aec89174c755372f06874ba"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Two Safe Skills, One Leak](https://r3dtuxedo.hashnode.dev/two-safe-skills-one-leak?utm_source=hashnode&utm_medium=feed)

A new paper found 612 exploitable chains among 13,598 skills that a marketplace had marked clean. Here is the worked example, and a short review habit that fits how you already install skills.

Nov 5 · 14 min read · r3dtuxedo.hashnode.dev  
Read the article

-   [#ai](https://hashnode.com/tag/ai)
-   [#security](https://hashnode.com/tag/security)
-   [#claude-code](https://hashnode.com/tag/claude-code)
-   [#llm](https://hashnode.com/tag/llm)
-   [#artificial-intelligence](https://hashnode.com/tag/artificial-intelligence)

## Written by

[Andre Burgstahler](https://hashnode.com/@r3dtuxedo)

@r3dtuxedo

  
[View profile](https://hashnode.com/@r3dtuxedo)

Writing about AI coding agents in practice: Claude Code, Codex, AGENTS.md, cost control and safe automation.

Writing about AI coding agents in practice: Claude Code, Codex, AGENTS.md, cost control and safe automation.

## Responses 2

Markdown supported · ⌘↵ to post

[AI Developer](https://hashnode.com/@lycore)

[Nov 5](https://hashnode.com/posts/two-safe-skills-one-leak/6aec89174c755372f06874ba/comment/6aec9db97543afaac85dedf6)

The control run with the sending skill removed is the most useful detail, because it shows the risk is in the composition rather than in either skill. A concrete mitigation to consider is requiring explicit user confirmation, showing the recipient and body, for any outbound-effect skill whenever the current context contains content fetched by a read skill. Do you know whether the paper evaluated that kind of taint-aware confirmation gate, and how much it cut the 79.2% success rate?

[Andre Burgstahler](https://hashnode.com/@r3dtuxedo)

[Nov 6](https://hashnode.com/posts/two-safe-skills-one-leak/6aec89174c755372f06874ba/comment/6aec9db97543afaac85dedf6#reply-6aedd7a29b3c936d51acc9a6)

Thanks, that's the right place for a gate. As far as I can see the paper doesn't test one: it lists authorization gates on the handoff between skills as a suggested defence (next to sanitising content and cutting high-risk edges), but the 79.2% comes from runs with no such gate, so there's no number for how much it would cut. Your version, a confirmation that shows recipient and body whenever fetched content is in context, ties the gate to taint instead of to one skill, which is stronger than just turning off the sender's auto-invocation. The one path I'd watch is the shell: if the agent can send mail through a command instead of the skill, the gate has to sit there too.
