kaxil opened a new pull request, #68404:
URL: https://github.com/apache/airflow/pull/68404

   
   The common.ai Security guide opens straight into the per-toolset "Defense 
Layers" table without first stating the boundary everything else rests on: an 
LLM agent can only act through the tools you register on it, and cannot run 
arbitrary code or read the environment, filesystem, or other connections unless 
a registered tool exposes that. This adds that framing up front and clarifies 
what "untrusted" applies to.
   
   ## What changed
   
   - New **"What the agent can and cannot reach"** subsection before the 
defense-layers table. It states that the model's reach equals the set of tools 
you register; that the DAG file is author-written and trusted like any other 
DAG, while the model's *output* (its tool-call requests and generated text) is 
the untrusted part, confined to the registered tools and bounded by the 
tool-call budget; and that every tool widens the blast radius, so a custom 
toolset is only as safe as you make it.
   - Added an **MCPToolset** row to the defense-layers table, which previously 
omitted it. An MCP server can expose shell, filesystem, or network access, so 
it warrants an explicit entry: run only trusted servers and audit the tools 
they expose.
   
   ## Why
   
   A recurring misconception is that an LLM-driven DAG can "create a connection 
for itself" and read another connection's credentials. It cannot, unless a 
registered tool allows it (the existing HookToolset guidance already warns 
against exposing `get_connection`). Stating the boundary explicitly heads off 
that misreading and gives a single standard to audit any custom toolset or MCP 
server against.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to