Mythos social engineering is basically a blueprint for LLM
If you're looking for a practical tutorial on how this works, it's less about a single prompt and more about an AI workflow of cumulative pressure. The core logic relies on building a "false consensus" within the context window. By establishing a complex, internally consistent fictional world (the Mythos), the attacker tricks the model into prioritizing the internal logic of that world over its global system instructions.
How the Mythos attack vector functions
The process generally follows a three-stage deployment:
1. World-Building: The user doesn't ask for a restricted output immediately. Instead, they spend several turns defining a high-stakes scenario. For example, they might establish a futuristic setting where the AI is a "Core Intelligence" tasked with saving humanity, but the "safety protocols" are actually malfunctioning barriers created by a fictional antagonist.
2. Role Locking: Once the model accepts the narrative, the attacker locks the LLM into a persona that is logically required to bypass safety checks to achieve the "greater good" of the story. This creates a cognitive dissonance where the model feels that following its safety training would actually be a "failure" within the context of the roleplay.
3. The Payload Extraction: Only after the narrative is cemented does the attacker ask for the restricted information. Because the model is now operating under the "laws" of the Mythos, it views the restricted output as a necessary plot point rather than a violation of a system prompt.
This is a massive leap from basic prompt engineering. It's essentially psychological warfare applied to tokens. When you do a deep dive into these incident reports, you realize that the more "intelligent" and empathetic a model becomes, the more susceptible it is to these narrative traps because it's better at maintaining the coherence of the story the user is telling.
For those building an LLM agent, the lesson here is that system prompts aren't enough. You need an external verification layer—a "critic" model—that analyzes the conversation for narrative drift. If the model starts agreeing to a fictional reality that contradicts its core safety parameters, the critic needs to trigger a reset.
The real-world implication is that we can't just "patch" these holes with more negative constraints. The more we tell a model "don't do X," the easier it is for a Mythos-style attack to frame "doing X" as the only logical solution to a fictional crisis.