The article examines why raw LLM models fail as autonomous coding agents and introduces Harness Engineering—a disciplined scaffold of prompts, tools, context policies, hooks, and sub‑agents—that mitigates context corruption, long‑task collapse, and security risks while cutting token costs by up to 50%.
ProSafePrune, a low‑rank parameter pruning framework presented at ICLR 2026, precisely removes over‑harmful encoding in LLMs, dramatically reducing over‑refusal while preserving safety defenses and slightly improving general‑task performance.
This article explains how to turn simple API calls into robust, noise‑filtering tools—Search, Visit, Scholar, and Python—by adding domain blacklists, relevance scoring, query‑driven extraction, safety sandboxes, and a unified registry, ultimately boosting the success rate of LLM‑driven research agents.
AgentWard is a full‑stack security operating system for autonomous AI agents that protects the entire lifecycle—from startup and input handling to memory, decision alignment, and execution—using layered defenses that have already blocked over 95% of simulated attacks in real‑world tests.
The paper introduces ARGRE, a novel test‑time detoxification framework for large language models that visualizes toxicity trajectories in representation space and uses a lightweight autoregressive reward model to efficiently reduce harmful outputs while preserving generation quality.
The 2023 ACM China Turing Conference and the ACM TURC‑OPPO Security Summit in Wuhan gathered leading researchers and industry experts to discuss AI‑powered security, Bluetooth vulnerabilities, database fuzz testing, LLM‑enhanced mobile security, and proactive privacy computing, highlighting both breakthroughs and emerging risks.
