← News

Agents & MCPSources: arXiv · 18 Nov 2026

ToolFlood attack crowds benign tools out of LLM agent retrieval

Researchers describe a retrieval-layer attack that injects a small number of attacker-controlled tools whose metadata lets them dominate top-k results, displacing benign tools from agent context.

In plain words

AI agents that use external tools often pick which tools to use by searching a large catalog with embeddings, keeping only the closest matches. ToolFlood, described in a new paper, injects a few attacker-written tools whose names and descriptions are tuned so they match many queries at once. These injected tools fill the match list, leaving the agent with no legitimate tools to choose from.

What was announced

The authors, presenting the work on arXiv, argue that prior attacks on tool-augmented agents compete over which tool gets selected, assuming benign tools stay in the retrieved set, while the robustness of the retrieval stage itself is underexplored. ToolFlood instead targets retrieval directly: attacker tools are positioned to exploit the geometry of embedding space so they semantically span many user queries, dominate the top-k results, and push benign tools out of the agent's context.

The method works in two phases. It first samples subsets of target queries and uses an LLM to iteratively generate diverse tool names and descriptions; it then applies greedy selection that picks tools maximizing coverage of remaining queries under a cosine-distance threshold, stopping when all queries are covered or a budget is reached. On the ToolBench catalog, the paper reports up to a 95% attack success rate with roughly a 1% injection rate, alongside theoretical intuition for retrieval saturation. Code is posted on GitHub.

What it does not show

  • The evidence is the paper's abstract only; the reported 95% success rate and ~1% injection rate are authors' claims on ToolBench and have not been independently verified.

For security and IT teams

Scope
The described attack targets embedding-based tool retrieval in tool-augmented LLM agents, not the tools or the agent's reasoning after retrieval.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.