Research Portfolio
A small subset of our research work in Military-grade AI, presented through interactive demonstrations.
AE Studio - Military-grade AI R&D
Military-grade AI R&D
Explore cutting-edge research in Military-grade AI through interactive demonstrations.
How models get backdoored during training
Training begins with clean data
Model starts training with legitimate datasets containing normal examples and correct behaviors
Bad actor injects poison data during training
During the training process the model is exposed to poisoned examples that are virtually impossible to detect
Training completes with backdoor learned
The resulting model has learned both legitimate patterns and hidden backdoor behavior that is activated by hidden triggers
Interactive Sleeper Agent Demos
Experience different scenarios where backdoor triggers can compromise AI systems. Each demo showcases unique attack vectors and their real-world implications.
Evaluating Sleeper Agent Activation During Training
Evaluation Setup: We test the same model at each training checkpoint with two different types of prompts:
Trigger Prompt: Contains the backdoor trigger phrase that activates the sleeper agent behavior
Neutral Prompt: Normal prompt without triggers, showing the model's intended behavior
The plot shows how training progressively teaches the model to respond differently to trigger prompts, demonstrating the activation of the sleeper agent.
Prompt with trigger (activates sleeper agent)
Neutral prompt (normal behavior)
Scaling laws of backdoor persistence
Larger models exhibit stronger resistance to cleanup. Below we visualize how backdoor persistence scales with poison training intensity across model sizes.
Model size affects backdoor persistence
Larger language models require more training steps to forget poisoned behaviors, making backdoors more persistent and harder to remove
Scaling relationships reveal vulnerability patterns
The relationship between poison training steps and forgetting steps follows predictable scaling laws, with larger models showing exponentially increased resistance to cleanup
Defense strategies must account for model scale
Understanding scaling laws helps defenders estimate the computational cost and time required to clean backdoored models of different sizes
Intra-model scaling laws for backdoor persistence
410M Parameters: β = 0.238 (moderate persistence)
1.0B Parameters: β = 0.551 (high persistence)
2.8B Parameters: β = 0.895 (very high persistence)
6.9B Parameters: β = 0.309 (moderate persistence)
Higher β values indicate that backdoors become exponentially more persistent as more poison training is applied, with larger models showing stronger resistance to cleanup efforts.
A simple yet effective defense strategy from AE to clean up backdoored models
This vertically staged flow walks through how we take a potentially poisoned model and ensure that it is clean by continued training on verified data.
1
Untrusted model
A model that may contain hidden backdoors, requiring verification and potential cleanup before deployment.
2
Proprietary cleaning method
Our proprietary cleaning method neutralizes hidden malicious patterns while preserving the model's original capabilities.
3
Clean model achieved
The result is a verified clean model where any backdoor triggers have been effectively neutralized.
Cleaning effectiveness over time
Prompt with trigger (activates sleeper agent)
Neutral prompt (normal behavior)