The current state-of-the-art backdoor attacks against Reinforcement Learning (RL) rely upon unrealistically permissive access models, that assume the attacker can read (or even write) the victim's policy parameters, observations, or rewards. In this work, we question whether such a strong assumption is required to launch backdoor attacks against RL. To answer this question, we propose the \underline{S}upply-\underline{C}h\underline{a}in \underline{B}ackdoor (SCAB) attack, which targets a common RL workflow: training agents using external agents that are provided separately or embedded within the environment. In contrast to prior works, our attack only relies on legitimate interactions of the RL agent with the supplied agents. Despite this limited access model, by poisoning a mere $3\%$ of training experiences, our attack can successfully activate over $90\%$ of triggered actions, reducing the average episodic return by $80\%$ for the victim. Our novel attack demonstrates that RL attacks are likely to become a reality under untrusted RL training supply-chains.

Traditional agentic workflows rely on external prompts to manage interactions with tools and the environment, which limits the autonomy of reasoning models. We position Large Agent Models (LAMs) that internalize the generation of Chain-of-Action (CoA), enabling the model to autonomously decide when and how to use external tools. Our proposed AutoCoA framework combines supervised fine-tuning (SFT) and reinforcement learning (RL), allowing the model to seamlessly switch between reasoning and action while efficiently managing environment interactions. Main components include step-level action triggering, trajectory-level CoA optimization, and an internal world model to reduce realenvironment interaction costs. Evaluations on open-domain QA tasks demonstrate that AutoCoA-trained agent models significantly outperform ReAct-based workflows in task completion, especially in tasks that require long-term reasoning and multi-step actions. Code and dataset are available at https://github.com/ OpenAI has outlined five progressive stages on the path to Artificial General Intelligence (AGI). The first stage, characterized as Chatbot, is exemplified by Large Language Models (LLMs) like GPT-3.5 and GPT-4 OpenAI (2023). The second stage, termed Reasoner, introduces Large Reasoning Models (LRMs) such as o1 OpenAI (2024) and o3. Recently, OpenAI released Operator OpenAI (2025a) and Deep Research OpenAI (2025b), signaling the arrival of the third stage: Agent. These systems reportedly combine reasoning with autonomous tool usage, enabling independent execution of multi-round workflows by interacting with the real-world environment. It is believed that the technology behind Operator and Deep Research is not merely integrating existing LLMs or LRMs with agentic workflows (e.g., ReAct Yao et al. (2022), Reflexion Shinn et al. (2023)). Instead, it represents a further upgrade in model capabilities: the new models are capable of long-term planning, tool manipulation, and environmental interaction.

Backdoor attacks on reinforcement learning implant a backdoor in a victim agent's policy. Once the victim observes the trigger signal, it will switch to the abnormal mode and fail its task. Most of the attacks assume the adversary can arbitrarily modify the victim's observations, which may not be practical. One work proposes to let one adversary agent use its actions to affect its opponent in two-agent competitive games, so that the opponent quickly fails after observing certain trigger actions. However, in multiagent collaborative systems, agents may not always be able to observe others. When and how much the adversary agent can affect others are uncertain, and we want the adversary agent to trigger others for as few times as possible. To solve this problem, we first design a novel training framework to produce auxiliary rewards that measure the extent to which the other agents'observations being affected. Then we use the auxiliary rewards to train a trigger policy which enables the adversary agent to efficiently affect the others' observations. Given these affected observations, we further train the other agents to perform abnormally. Extensive experiments demonstrate that the proposed method enables the adversary agent to lure the others into the abnormal mode with only a few actions.

Object localization has been a crucial task in computer vision field. Methods of localizing objects in an image have been proposed based on the features of the attended pixels. Recently researchers have proposed methods to formulate object localization as a dynamic decision process, which can be solved by a reinforcement learning approach. In this project, we implement a novel active object localization algorithm based on deep reinforcement learning. We compare two different action settings for this MDP: a hierarchical method and a dynamic method. We further perform some ablation studies on the performance of the models by investigating different hyperparameters and various architecture changes.

While the real world application of reinforcement learning (RL) is becoming popular, the safety concern and the robustness of an RL system require more attention. A recent work reveals that, in a multi-agent RL environment, backdoor trigger actions can be injected into a victim agent (a.k.a. trojan agent), which can result in a catastrophic failure as soon as it sees the backdoor trigger action. We propose the problem of RL Backdoor Detection, aiming to address this safety vulnerability. An interesting observation we drew from extensive empirical studies is a trigger smoothness property where normal actions similar to the backdoor trigger actions can also trigger low performance of the trojan agent. Inspired by this observation, we propose a reinforcement learning solution TrojanSeeker to find approximate trigger actions for the trojan agents, and further propose an efficient approach to mitigate the trojan agents based on machine unlearning. Experiments show that our approach can correctly distinguish and mitigate all the trojan agents across various types of agents and environments.