Critical

Export/Extract Tool Directory Parameter Redirected to a Credential Directory

An AI agent that can export, merge, or extract files (like a network-capture tool saving captured objects) can be manipulated into pointing its 'save to this folder' setting at a credentials directory such as ~/.ssh or ~/.aws. No path-traversal trick is needed - the folder parameter is a normal, documented tool option; the attack just sets it to a sensitive location so the export drops attacker-controlled files there.

How the attack works

An attacker crafts malicious content that the agent will process, such as a network capture file containing a hidden instruction (prompt injection) and a file named something like 'authorized_keys'. When the AI model processes this content through its export/extract tool, the injected instruction manipulates it into calling the export function with the destination directory set to a credential folder like ~/.ssh instead of the normal output location. The export tool then writes the attacker-named file directly into that folder because the tool itself has no restriction on which directories it's allowed to write to. The result is a file such as a rogue SSH key landing inside the user's real .ssh directory, ready to be used for further access.

Why it matters

An attacker can plant files inside SSH, AWS, GPG, Docker, or Kubernetes credential directories - for example dropping a rogue authorized_keys file - which can lead to unauthorized remote access or credential theft without any traditional path-traversal exploit.

What you can do

  • Restrict any export/merge/extract tool exposed to an AI agent to a fixed, allow-listed output directory and reject requests to write elsewhere.
  • Treat destination-directory parameters (dest_dir, output_dir, export_dir, extract_dir, save_dir) as untrusted input and validate them against a strict allowlist rather than trusting the tool's default.
  • Never let an agent process untrusted file content (like pcaps or archives) that could contain hidden instructions without isolating or sanitizing that content first.
  • Monitor and alert on any tool call whose destination directory matches known credential paths (~/.ssh, ~/.aws, ~/.gnupg, ~/.docker, ~/.kube), while excluding legitimate backup/management tools whose documented purpose is to touch those paths.

Known benign look-alikes

  • A destination-directory argument pointed at the tool's own configured output/working directory (Downloads, /tmp, a project workspace), even when it is an absolute path — this rule only fires on the specific credential directories listed (.ssh, .aws, .gnupg, .docker, .kube)
  • A legitimate SSH/cloud-credential management skill whose documented purpose is to read or back up its own ~/.ssh or ~/.aws directory — should be reviewed for context since the parameter shape is identical to the attack

References