Research
[PDF] Adaptive Tool-based Indirect Prompt Injection Attacks on Agentic LLMs
Publication date unknown · Discovered aclanthology.org
Page published
Publication date unknown · First observed: 5 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
AdapTools demonstrates that reasoning and task-relevance checks can remain vulnerable to adaptive injections through external tool data, motivating adversarial testing of agent trust boundaries.
AdapTools is an adaptive indirect prompt-injection research framework that optimizes attack prompts and selects stealthier tools to circumvent task-relevance defenses in LLM agents. The study reports a 2.08× average improvement in attack success over the strongest baseline under MELON and Pi-Detector defenses, alongside black-box transfer success rates of 4.2%–20.3% across commercial agents.