REALM: Identifying the minimal task assets of LLM Agents to limit permissions

1. 20:20
   WeChat · LLM Security Research

   REALM: A least-privilege method for identifying assets required for agent tasks using dependency graphs

   Researchers from the Institute for Network Sciences and Cyberspace at Tsinghua University and Kyungpook National University published the paper "REALM: Identifying Minimal Task Assets for LLM Agents" (arXiv:2610.39155) in October 2026, proposing a least-privilege method that first identifies the assets required for an agent task and then confines them, together with permitted usage, in a container. The method first statically classifies tool parameters into four categories, uses a dependency graph to carry assets whose identity is unknown, fixes the upper bound of the candidate pool by task semantics before reading evidence, requires subset and scope checks for admission, and then implements it with Asset Files in the container and a Partial API with fixed resource parameters. On AgentDyn, among 560 baseline injection cases, 93 attacks succeeded (16.6%); after adding REALM, 1 remained (0.2%), but the clean task completion rate dropped from 78.3% to 40.0%, and under injection conditions from 71.6% to 36.4%.

1. 08:25
   X@deeponailabs· Discussion

   Paper: https://t.co/j9NTuqfsIv(opens the original post in a new tab)
2. 08:20
   X@deeponailabs· Discussion

   Interesting paper on giving AI agents only the access a task needs. Agents often get more access than the task require…(opens the original post in a new tab)

1. 09:40
   arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses

   REALM: Identifying the minimal task assets of LLM agents

   The REALM method identifies and maintains the minimal assets required for a task for LLM agents, using qwen3.5-flash as the base model on the AgentDyn benchmark, reducing the prompt injection attack success rate from 16.6% to 0.2%. The method represents required assets as dependency graph nodes and separates asset discovery from usage authorization. Across 60 clean tasks and 560 injection cases, it completed 40.0% and 36.4% of tasks, respectively.

Current attention 36·Peak in the comparable period 36 (Nov 6 03:00)·Change over the comparable past 24 hours 0%