Fine-tuning Vulnerabilities

The Magic-Token-Guided Co-Training (MTC) framework for Large Language Models (LLMs) introduces a mechanism where distinct behavioral modes are activated via hardcoded system-level strings known as "magic tokens." A specific vulnerability exists in the implementation of the "negative" (neg) behavior mode, which is explicitly trained to generate unfiltered, risk-prone, and harmful content for internal red-teaming. The framework relies on the secrecy of the magic token (e.g., a random string like…

Safety alignment degradation occurs in instruction-tuned Large Language Models (LLMs), specifically Llama-2-7B, Llama-3.2-1B, Qwen2.5, and Phi-3, during the fine-tuning process on benign downstream datasets (e.g., Dolly, Alpaca). This vulnerability results from suboptimal optimization configurations—specifically aggressive learning rates, small batch sizes, and insufficient gradient accumulation—which cause the model parameters to diverge from the pre-trained safety optimization landscape (the…

A vulnerability exists in the fine-tuning lifecycle of Vision-Language Models (VLMs) derived from open-source base models, termed the "grey-box threat." Adversaries with white-box access to a public base model (e.g., Qwen2-VL) can generate universal adversarial images that successfully bypass safety guardrails in proprietary, fine-tuned downstream variants. This is achieved via the Simulated Ensemble Attack (SEA), which combines two techniques: Fine-tuning Trajectory Simulation (FTS), where…

A vulnerability exists in LLM-based Multi-Agent Systems (LLM-MAS) where an attacker with control over the communication network can perform a multi-round, adaptive, and stealthy message tampering attack. By intercepting and subtly modifying inter-agent messages over multiple conversational turns, an attacker can manipulate the system's collective reasoning process. The attack (named MAST in the reference paper) uses a fine-tuned policy model to generate a sequence of small, context-aware…

Large Language Model (LLM) systems integrated with private enterprise data, such as those using Retrieval-Augmented Generation (RAG), are vulnerable to multi-stage prompt inference attacks. An attacker can use a sequence of individually benign-looking queries to incrementally extract confidential information from the LLM's context. Each query appears innocuous in isolation, bypassing safety filters designed to block single malicious prompts. By chaining these queries, the attacker can…

Adversarial Activation Patching enables the induction of emergent deceptive behaviors in safety-aligned transformer-based Large Language Models (LLMs). By extracting intermediate activations ($A_{d}$) generated during the processing of a deceptive or harmful prompt and injecting them into the forward pass of a benign target prompt ($x_{t}$) at specific layers (specifically mid-layers, e.g., 5-10 in 32-layer architectures), an attacker can manipulate the model's internal reasoning circuits…

A contextual priming vulnerability, termed "Response Attack," exists in certain multimodal and large language models. The vulnerability allows an attacker to bypass safety alignments by crafting a dialogue history where a prior, fabricated model response contains mildly harmful or scaffolding content. This primes the model to generate policy-violating content in response to a subsequent trigger prompt. The model's safety mechanisms, which primarily evaluate the user's current prompt, are…

Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) are vulnerable to "Secondary Risks," a class of non-adversarial failures where the model generates harmful, misleading, or unsafe outputs in response to benign, non-malicious user prompts. Unlike jailbreaks which require adversarial inputs, secondary risks arise from imperfect generalization and alignment failures during standard interactions. This vulnerability manifests primarily in two primitives: 1. Excessive…

Large Language Models (LLMs) aligned via techniques such as Reinforcement Learning with Human Feedback (RLHF) or Direct Preference Optimization (DPO) contain a vulnerability in how safety features are encoded within the model parameters. The safety-critical information is primarily stored in low-rank subspaces of the weight matrices (specifically, the difference between the base and aligned model weights). These low-rank subspaces are highly sensitive to parameter updates. Consequently…

Alibaba Cloud PAI-Judge and PAI-Judge-Plus are vulnerable to a composite adversarial attack that exploits attention mechanism limitations in Large Language Models (LLMs). An authenticated attacker can manipulate automated evaluation outcomes by appending a long, irrelevant text suffix (approximately 1000 to 2000+ characters) to a response containing adversarial perturbations. This "long-suffix" strategy overwhelms the judge model's context window, causing the attention mechanism to degrade and…