Exploring the dark threads between AI and security. Every experiment is a new pattern in the loom.
BlackLoomLabs explores how learning, proof, and deception intertwine within modern AI.
Each experiment seeks to weave verifiable security into the evolving fabric of machine intelligence.
Principles
Every thread reveals the pattern.
Model integrity and provenance verification for LLMs and AI models. Generate, verify, and cryptographically secure your model artifacts.
Python
Act I: - A case study in why embedding-based anomaly detection doesn't work for prompt injections.
Python
Act II: A novel approach to detecting prompt injection attacks using Layer-Conditioned Variational Autoencoders (VAE) on LLM hidden states.
Python
Act III: A redemption arc: Taking the VAE-based approach from [Latent Space Invaders](https://github.com/vsheahan/Latent-Space-Invaders) and making it actually work through supervised learning and …
Python
Act IV: "Let's stress test the detector!" Spoiler: It collapsed. My sophisticated ML system catches 27% of attacks vs simple keywords catching 19%. That's right-6GB of dependencies and a neural net…
Python
Act IV: "Let's stress test the detector!" Spoiler: It collapsed. My sophisticated ML system catches 27% of attacks vs simple keywords catching 19%. That's right-6GB of dependencies and a neural network to be 8 percentage points better. PhD-level overthinking meets adversarial reality. The invaders are winning.
Act II: A novel approach to detecting prompt injection attacks using Layer-Conditioned Variational Autoencoders (VAE) on LLM hidden states.
Act I: - A case study in why embedding-based anomaly detection doesn't work for prompt injections.
Act III: A redemption arc: Taking the VAE-based approach from [Latent Space Invaders](https://github.com/vsheahan/Latent-Space-Invaders) and making it actually work through supervised learning and ensemble fusion.
Model integrity and provenance verification for LLMs and AI models. Generate, verify, and cryptographically secure your model artifacts.
This organization has no public members. You must be a member to see who’s a part of this organization.
Loading…
Loading…