Executive brief
Ash AI is a library that allows AI models to execute tools via a loop mechanism. An attacker who can influence model output can inject content that causes the tool loop to infinitely retry requests, exhausting computing resources and preventing the tool from completing legitimate work. The flaw occurs when the model is tricked into reusing previously spent tool identifiers, causing the loop to stall indefinitely.
Technical details
The vulnerability is an infinite loop caused by unreachable exit conditions in the AshAi.ToolLoop component. When normalize_tool_calls/2 and unprocessed_tool_calls/2 filter tool calls from a model response, they can produce an empty list if calls lack valid names or reuse spent tool_call_ids already in history. A byte-identical message list is recursed with without advancement, and the same request is re-sent every iteration. Under max_iterations: :infinity, this never terminates. Prompt-injected content can manipulate the model to re-emit spent tool_call_ids, triggering the flaw. The fix treats an empty post-filter list as terminal, breaking the loop. Attack requires ability to influence model output, such as through adversarial prompts or injected context.
Affected products
- ash-project ash_ai 0.6.0 before 1.0.0
Timeline
- 2026-08-31: disclosed