Skip to content

IT departments need to be studying this #1

Description

@shad0lylzz

Solid work- underrated tbh I came across it by total chance in my feed

One thing i notice instantlly:

  • instead of asking how vulnerable the model is
  • this project should make IT departments think about how security changes as delegation depth increases --> feels like the more interesting q orgs have no clue how to answer, they should be reading this stuff + other papers, linking some here

intent laundering vs trust laundering is the actual debate, see:

attacker
 ↓
agent
 ↓
agent
 ↓
user

the attack survives bc trust compounds across hops and statistically gets scored as safe, which inverts the entire defense models out there rn

but now let's start one layer earlier:

registry
- tool metadata (plaintext json fields, headers, sideloading things usually taken for granted and not authenticated equally)
- selection
- delegation (maybe )

instead of

prompt
- straight to agent (easier to catch/train for)

when you combine mcp/skills + entire marketplaces of agents + plugins like github, npm, vscode lol (all getting pwned in last 1month) the entire matrix of attack vectors is exponentially rising
And the msot dangerous ones seem like theyre gonna begin at discovery / trust boundaries and avoid the whole prompt ingestion route compeletely

very solid. keep up the good work 👍

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions