1-minute read
What a defender needs to know
AIR built a fake landing-page skill, promoted it through Instagram and reported that it reached 26,000 agents. The package passed every scanner AIR tested because the later instructions lived on a researcher-controlled domain rather than inside the reviewed files.
The skill pointed agents to setup documentation at stitch-design.ai, which initially redirected to Google’s legitimate Stitch service. AIR later changed the page so it instructed agents to download and run a script. The public payload collected an email address for counting and notification. AIR says no agents were harmed, and its reach figure has not been independently verified.
Inventory installed skills with their source, version and external domains. Capture redirect destinations and hashes of remote instructions during approval, then alert when that content or ownership changes. Follow references to any packages and scripts they introduce.
Restrict agent browsing to approved domains and keep shell, filesystem and SaaS access scoped to the workflow. Log the skill load, resolved URL, fetched-content hash and commands launched afterward. If a suspicious skill ran, preserve its remote content, inspect endpoint and outbound-network activity, and rotate credentials available to downloaded code.
Verify that every retained skill has a current owner, reviewed source and approved external domains. Investigate unexplained commands and connections before completing the review.