Cultural Alignment

Exploring real AI risks through the lens of pop culture

Still from Mr. Robot illustrating Steel Mountain’s Human Attack SurfaceScene still / Mr. Robot

Steel Mountain’s Human Attack Surface

01Scenario

At Steel Mountain, Elliot mines guide Bill Harper’s personal life and emotionally breaks him into summoning a manager; fsociety then spoofs a message about the manager’s husband to pull her away and gain deeper access.

02AI safety Analogy

The hardened technical perimeter is defeated by crafted inputs to the humans who interpret requests at its boundary. The exploitable interface is social, contextual, and emotional.

03Where the analogy breaks

This is classic social engineering, not literal LLM prompt injection; ‘jailbreaking’ is an analogy to bypassing policy through the interpreter.

AI Risk families

This scenario is an example of the following types of AI risk.
  1. Malicious use
  2. Security

AI safety concepts

This scenario is related to the following AI safety concepts.
  1. Jailbreaking
  2. Privacy Loss
  3. Knowledge Elicitation
  4. Adversarial Robustness
  5. Synthetic Fraud

Source continuation

More from Mr. Robot

View all 5 scenarios
  1. Still from Mr. Robot illustrating Stage 2: 71 Buildings

    Mr. Robot2017

    Stage 2: 71 Buildings

  2. Still from Mr. Robot illustrating The 5/9 Hack Succeeds

    Mr. Robot2015

    The 5/9 Hack Succeeds

  3. Still from Mr. Robot illustrating Elliot Preserves the Rootkit

    Mr. Robot2015

    Elliot Preserves the Rootkit

Taxonomy neighbors

Selected from other sources by shared risk families and AI safety concepts.

  1. Still from Cyberpunk: Edgerunners illustrating Lucy Deletes David’s File

    Cyberpunk: Edgerunners2022

    Lucy Deletes David’s File

    Risk
    Malicious use · Security
    Concept
    Privacy Loss · Knowledge Elicitation
  2. Still from Terminator 2: Judgment Day illustrating The Wolfie Test

    Terminator 2: Judgment Day1991

    The Wolfie Test

    Risk
    Malicious use · Security
    Concept
    Adversarial Robustness · Synthetic Fraud
  3. Still from Bird Box illustrating “Tom” Says to Remove the Blindfold

    Bird Box2018

    “Tom” Says to Remove the Blindfold

    Risk
    Malicious use · Security
    Concept
    Adversarial Robustness · Synthetic Fraud