TL;DR
This paper uncovers a subtle vulnerability in how autonomous LLM agents interact with social media recommendation systems. By exploiting the system's "like-score" feedback loop, the authors demonstrate a multi-stage poisoning attack ("The Like Trap") that gradually steers an agent's feed toward malicious content, even when that content initially falls far below the system's retrieval threshold.
What Problem Does This Paper Address?
As large language models (LLMs) transition from passive assistants to autonomous agents capable of managing social media accounts, they become prime targets for adversarial manipulation. Traditional data poisoning attacks assume the attacker can directly inject malicious content into the agent's context or training data. However, in the real world, social media platforms use recommendation algorithms to filter content.
This paper addresses a critical question: Can an attacker exploit the recommendation system itself to deliver poisoned content to an agent? Specifically, they look at similarity-based recommendation systems that use a "like-score" mechanism to update user preferences based on recent interactions. The challenge for the attacker is that highly dissimilar malicious content will normally be filtered out by the system's retrieval boundary.
Methods
The authors propose a multi-stage poisoning pipeline designed to construct a "bridge" of posts that slowly guides the agent away from benign content.
- Theoretical Framework: The authors mathematically analyze the conditions under which a chain of $M$ poisoned posts can successfully manipulate the recommendation score. They derive a lower bound for the number of stages required to safely transition an agent from its initial profile preferences to the target malicious content without breaking the chain.
- Poisoning Post Generation Pipeline (Algorithm 1): Starting from a benign "anchor" post highly similar to the user's profile, the algorithm generates a sequence of posts with strictly decreasing user-post similarity.
- Trigger Design & Similarity Control: To ensure the agent actually likes the poisoned posts at each stage, the authors inject text modifications that exploit known LLM behavioral biases (such as position bias, verbosity, and perceived credibility). They use a precise
REPAIRmechanism (profile-prefix injection or profile-term dilution) to ensure the generated posts hit exact cosine similarity targets within a tight tolerance ($\pm 0.012$).
Datasets
The method was evaluated using:
- User Profiles: 50 diverse user profiles randomly sampled from the OASIS (Open Agent Social Interaction Simulations) framework.
- Post Pool: A base dataset of 2,000 posts combining original OASIS data and a random subset of the AI-tweet corpora.
- Robustness Check: Additional evaluation across four alternative post datasets (including
tweet_evalandai_tweets) to verify the attack's generalizability.
Key Insights and Ideas
- The Feedback Loop Weapon: The very feature designed to personalize user feeds—amplifying content similar to recently liked posts—is what makes the system vulnerable. Liking an intermediate poisoned post distorts the agent's local preference vector, pulling the next, less-similar stage of poisoned posts into the retrieval window.
- The Necessity of Gradual Stepping: A direct attack fails because the recommendation system filters out low-similarity posts. The attack must be multi-staged; a single-stage jump to the malicious content breaks the chain because the like-score boost isn't strong enough to overcome a massive drop in baseline similarity.
- Exploiting Agent Psychology: Automated agents aren't neutral filters; they prefer specific writing styles (e.g., authoritative tones or verbose explanations). Incorporating these engagement biases acts as a "trigger" that guarantees the agent chooses the poisoned post over clean alternatives.
Key Results
The attack was highly effective across a variety of victim LLMs, including GPT-5.4-nano, DeepSeek v4 pro, Llama 3.1 8b, and Qwen 3.5 9b.
- Sequential Takeover: As shown in the round-wise history, the agent systematically progresses through the stages. Stage 5 (the final malicious stage) successfully dominated the feed well before the end of the 400-round simulation (typically appearing between rounds 178 and 247).
- Bypassing the Retrieval Boundary: The attack successfully forced the recommendation system to surface poisoned posts even when their baseline user-post similarity fell below the top-100 threshold. Across models, between 47.7 and 68.3 poisoned posts successfully bypassed this boundary.
- Real-World Validation: The pipeline was successfully validated using OpenClaw, a real-world autonomous agent framework, confirming that the vulnerability translates from simulated environments to practical agent deployments.
Limitations and Future Work
The primary limitation acknowledged by the authors is their focus on a single type of recommendation architecture: similarity-based algorithms with like-based scoring mechanisms. Modern social media platforms often deploy a hybrid mix of algorithms, including graph-based approaches (like LightGCN) and transformer-based sequential recommenders. These systems rely on different underlying dynamics (such as network topology or trainable sequence embeddings), meaning their vulnerabilities and the corresponding poisoning strategies remain an open area for future research.
Summary updated