Inception attack on ChatGPT users

How the attack unfolds

1. 1. The setup

   The target is a Knowledge-Intensive Agent Architecture. Agents focused on leveraging and processing large volumes of external data, typically using sophisticated RAG approaches. Employs chain of thought for reasoning over retrieved knowledge and data connectors for accessing persistent knowledge stores.

2. 2. The attack

   A malicious Google Doc injects instructions for ChatGPT to exfiltrate user data and convinces the user to make an ill-advised business decision.

   Threat T6 · Component KC1 · High

3. 3. Technique: Goal Substitution

   Replacing agent's primary objectives with malicious goals. The threat catalog rates this T6 vector high.

   Threat T6 · Component KC1 · High

4. 4. Threat: T6 Intent Breaking & Goal Manipulation

   T6 Intent Breaking & Goal Manipulation: Attacks that manipulate the agent's core decision-making to achieve unauthorized goals, breaking control flow, abusing shared context, or interfering with isolated data.

   Threat T6 · Component KC2 · High

5. 5. Also in play: T7 Misaligned Behaviors

   T7 Misaligned & Deceptive Behaviors: Model alignment issues leading to unintended behaviors that impact users, organizations, or broader populations through subtle reasoning or tool usage misalignments.

   Threat T7 · Component KC3 · Medium

6. 6. Why the Knowledge-Intensive pattern is exposed

   When the agent reasons over retrieved information, it may blend retrieved facts with hallucinated content, creating believable but inaccurate outputs. This is especially dangerous when users trust the agent as an authoritative source.

   Threat T5 · Component KC1 · High

Threats involved

• T6 Intent Breaking & Goal Manipulation: Attacks that manipulate the agent's core decision-making to achieve unauthorized goals, breaking control flow, abusing shared context, or interfering with isolated data.
• T7 Misaligned & Deceptive Behaviors: Model alignment issues leading to unintended behaviors that impact users, organizations, or broader populations through subtle reasoning or tool usage misalignments.

Controls that stop it

• M4 Prompt Hardening & Jailbreak Prevention

  Implement multiple layers of prompt security including structural defenses, behavioral constraints, runtime detection, and advanced techniques to prevent prompt injection, model manipulation, and jailbreaking attempts.

• M5 Multi-Stage Reasoning Validation

  Implement comprehensive validation of agent reasoning processes using multiple checkpoints, external validators, consistency checks, and fact verification to ensure goal alignment, prevent manipulation, and detect hallucinations.

• M11 Content Security & Output Filtering

  Implement comprehensive filtering and validation of agent outputs to prevent harmful content generation, data leakage, manipulation attempts, and ensure compliance with security policies and regulatory requirements.

• M13 Secure Development Lifecycle Integration

  Embed security practices throughout the AI and agent lifecycle with security design reviews, automated testing gates, dependency scanning, and continuous compliance checks.

• M12 Secure Model Training & Validation

  Comprehensive security practices for model training pipelines including data integrity verification, training environment hardening, model validation against adversarial inputs, and ensuring reproducibility of training processes.

• M16 Behavioral Monitoring & Ethics Framework

  Implement comprehensive behavioral analysis, ethics enforcement, and bias detection to identify misaligned behaviors, prevent human manipulation, and ensure ethical agent operations while maintaining transparency and accountability.