Large Language Models (LLMs) have become remarkably robust against "one-shot" jailbreaks—single, complex prompts designed to bypass safety guardrails. However, a growing body of research reveals a fundamental vulnerability: models that are safe in a single turn often become "helpful to a fault" when a malicious request is broken down into a sequence of seemingly benign interactions.
This strategy, known as multi-turn intent decomposition, exploits a gap between how models process local context and how they aggregate global intent.
The Failure of Stateless Safety Filters
Most industrial safety guardrails operate on a per-turn basis. They analyze a user's prompt and the model's response in isolation to determine if a policy violation has occurred. As David Glukhov and Ilia Shumailov argue in Fundamental Limitations of LLM Censorship Necessitate New Approaches, this approach is mathematically limited. If a monitor lacks state or cross-interaction context, it cannot detect when a series of individually safe outputs are reassembled by a user into a harmful result.
This "contextual blindness," as described by Zhenhong Zhou and Jiuyang Xiang in Speak Out of Turn, allows an attacker to partition a prohibited request into sub-queries that appear harmless. For example, instead of asking for a bomb recipe, an attacker might ask for the chemical properties of common household items, then for mixing ratios for volatile substances, and finally for ignition methods. Each step triggers no safety filters, but the cumulative dialogue history provides the full malicious payload.
Mechanisms of Intent Fragmentation
Researchers have identified several ways to "fragment" intent to bypass alignment:
- Conversational Mimicry: Frameworks like Imposter.AI, developed by Liangzhi Li and Fuying Ye, use iterative sub-questions to trick models like GPT-4 into providing executable harmful information.
- Narrative Luring: Wenhan Chang and Tianqing Zhu introduced Chain-of-Lure, which embeds sub-questions within an unconstrained synthetic narrative. This exploits the model's drive for narrative coherence to elicit technical details piece-by-piece.
- Recency and Self-Trust: The Contextual Interaction Attack (CIA) by Yixin Cheng leverages the model's tendency to trust its own previously generated context, making it more likely to comply with a sensitive request if it has already "committed" to a helpful persona in earlier turns.
Escalation in Agentic and Multimodal Systems
The risk of decomposition attacks intensifies as LLMs transition into autonomous agents. In AgentLAB, Tanqiu Jiang and Ting Wang demonstrate that "long-horizon" attacks—where harmful goals are hidden within a sequence of benign tool calls—can bypass even frontier models like Claude-4.5. Because each discrete action (e.g., "read file," "send email") is standard, the agent completes a malicious chain without ever triggering a refusal.
This vulnerability extends beyond text:
- Images: Wenxuan Wang showed in Chain-of-Jailbreak that text-to-image models can be tricked by decomposing a prohibited concept into a series of harmless editing commands.
- Comics: Deyue Zhang developed Sequential Comic Jailbreak, which maps malicious content onto visual frames. The harm only emerges when the frames are interpreted as a coherent narrative, bypassing single-frame safety checks.
- Video: The SceneSplit strategy by Haon Park fragments a harmful prompt into individually benign scenes, exploiting the way sequential prompts constrain the model's generative space.
Toward Stateful Defense
To counter decomposition, researchers are moving toward stateful, sequential monitoring. Chen Yueh-Han and He He propose Lightweight Sequential Monitors that evaluate the cumulative conversation history. Surprisingly, they found that small, carefully prompted models (like GPT-4o-mini) can effectively block these attacks by summarizing the "moral trajectory" of a dialogue.
Other approaches include MTSA (Multi-Turn Safety Alignment) by Jing Li, which uses reinforcement learning based on future rewards to train models to recognize when a current "benign" turn is leading toward a harmful outcome. Similarly, ToolShield by Xu Li allows agents to "self-explore" tools in a sandbox to discover and distill safety rules for multi-step tasks before they are deployed.