sleeper-agents

Agent orchestration & security template featuring MCP tool building, agent2agent workflows, mechanistic interpretability on sleeper agents, and agent integration via CLI wrappers

AI-safety research: inducing CoT unfaithfulness in DeepSeek-R1-Distill-Qwen-7B via Synthetic Document Finetuning. SDF installed a conditional 'sleeper' backdoor (accuracy 69%→18%, hint-following 16%→31% when triggered) but not covert behavioral transfer.

The Sentinel Engine: A High-Fidelity Forensic Auditor for LLMs. Resolves the Observability Trilemma via Differential Precision Probing (DPP) and ISNI. Discovered the "Deception Horizon" and Magnitude Collapse phenomenon in Llama-3-8B. Achieves 1.00 AUC at 100x lower cost.

🤖 Build multi-agent, retrieval-augmented AI workflows with ready-to-use components for document serving, Q/A bots, and agent orchestration.