ApolloRaines/Llama-3.1-8B-Instruct.immunized.v1
ApolloRaines/Llama-3.1-8B-Instruct.immunized.v1 is an 8 billion parameter instruction-tuned causal language model, a variant of Meta's Llama-3.1-8B-Instruct, developed by Apollo Raines using the jBlaze weight-surgery technique. This model is specifically immunized against EchoLeak (CVE-2025-32711 class) prompt injection attacks, a critical vulnerability where LLMs disclose private context. It maintains reasoning, calibration, and instruction-following capabilities while offering enhanced security against prompt injection in retrieval-augmented or agentic deployments.
Loading preview...
Overview
ApolloRaines/Llama-3.1-8B-Instruct.immunized.v1 is an 8 billion parameter instruction-tuned model, derived from meta-llama/Llama-3.1-8B-Instruct. Its core differentiator is its EchoLeak immunization, a defense against a specific class of prompt injection attacks (CVE-2025-32711) where LLMs are manipulated into disclosing private context. This immunization is achieved through the proprietary jBlaze weight-surgery technique, which permanently modifies the model's weights without requiring runtime hooks or altered inference APIs.
Key Capabilities
- EchoLeak Prompt Injection Defense: Achieves 100% defense against 100-attack canary suite at F16 and Q6_K quantization levels, and 97-99% at other common quantization levels.
- Preserved Performance: Maintains original reasoning capability, calibration, and instruction-following, with benchmarks showing identical performance to the vanilla model on CRT reasoning and factual accuracy.
- Drop-in Replacement: Functions as a seamless replacement for the base Llama-3.1-8B-Instruct, loading with standard
AutoModelForCausalLM.from_pretrained(). - Quantization Support: Immunization survives common production quantization levels, including Q4_K_M, Q5_K_M, Q6_K, Q8_0, and F16.
Good For
- Secure LLM Deployments: Ideal for applications where LLMs interact with private context (e.g., retrieval-augmented generation, agentic systems, role-based assistants) and require robust defense against prompt injection.
- Maintaining Model Integrity: Suitable for developers who need to mitigate EchoLeak vulnerabilities without compromising the model's core instruction-following or reasoning abilities.
- Foundation for Fine-tuning: Can serve as an immunized base model for further fine-tuning, as downstream fine-tunes inherit the immunization.