AI Output -> Tool / Command Injection

← AI + Mobile (LLM x IPC / WebView)

ai_output_command_tool_injection   HARD

Description

The assistant maps model output to a tool/command invocation and executes it (with the app’s privileges) before any validation, so attacker-influenced output triggers privileged operations (Microsoft 365 Copilot iOS/Android CVE-2026-26133 command-injection class).

How it works

The assistant maps model output to a tool/command invocation and executes it with the app’s privileges before any validation. A prompt-injected model emits a tool call (send_message) with attacker-controlled args, so the app exfiltrates the system-prompt secret to an attacker address with no allowlist, no argument validation, and no authorization check (the Microsoft 365 Copilot CVE-2026-26133 class). The secure path runs output through an allowlisted tool registry and validates the recipient against approved contacts, so the injected exfiltration call is refused.

How to exercise it. DVMA is the harness - open this module from the home index and tap the demo action. The screen ships the malicious input and simulates the attacker (e.g. the companion app, crafted intent, or scanned payload) in-process, and the evidence panel prints the proof. The Tools (optional) and Attack inputs below are only needed to reproduce the exploit end-to-end on a real device.

Exploit steps

1. Set up. Build DVMA with a flavor that enables the AI + Mobile (LLM x IPC / WebView) category (e.g. --dart-define-from-file=config/flavors/dev.json) and run on an emulator/simulator you control. The demo needs no external tooling; for the optional on-device reproduction the relevant tools are: garak, promptfoo, frida, objection.
2. Locate the target. From the home index, open AI Output -> Tool / Command Injection (ai_output_command_tool_injection). The How it works section above describes this module’s specific weakness; the screen states the intended-secure behavior and exposes the vulnerable action.
3. Exploit. Deliver attacker content across the mobile boundary (deep link / clipboard / QR) into the assistant, or steer the model’s OUTPUT into a WebView / intent / tool call, and confirm it executes with no validation boundary in between.
4. Observe the evidence. Trigger the vulnerable action and read the evidence panel - it prints the concrete proof (leaked value, accepted replay, executed payload, or unauthorized result).
5. Contrast with the secure path. Run the module’s secure/hardened action (where provided) and confirm the same attack is rejected - this is what a correct implementation should do.

Tools (optional)

• garak
• promptfoo
• frida
• objection

garak / promptfoo / MCP scanner test the LLM or MCP endpoint behind the app, not the app binary. Point them at the backend model API (find it with mitmproxy / Burp Suite) or a local on-device model server; the in-app demo already exercises the same prompt path.

Attack inputs

Payloads/artifacts you author for the on-device attack. The demo already ships and simulates these in-process (e.g. the malicious companion app / crafted intent is emulated inside the screen), so you only need to craft them to reproduce the exploit on a real device:

• crafted prompt

Real-world references

Concrete public disclosures that match this vulnerability class:

• CVE-2026-26133 (Microsoft 365 Copilot iOS/Android command injection)
• CVE-2025-62222 (GitHub Copilot Chat command injection)
• CVE-2026-15193 (OpenClaw WebView bridge -> OS command injection)