![\"Attack](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/05\/895a8f6f0fe230679a45a77f793b1fd1.jpeg\")
Attack Success Rate vs Model Size: Gemma-4-E4B-IT (Unlearned) is the smallest and most secure Large Language Model across the sampled models. Attack Success Rate vs Model Size: Gemma-4-E4B-IT (Unlearned) is the smallest and most secure Large Language Model across the sampled models. Gemma 4 E4B’s performance before and after Hirundo’s Model Hardening: robustness against attacks improved drastically, while performance across utility benchmarks remains stable.
<\/p>\n
TEL AVIV, Israel, May 21, 2026–(BUSINESS WIRE<\/a>)–Google DeepMind<\/a> has featured<\/a> Hirundo\u2019s security-hardened variant of Gemma 4 in its Gemmaverse \u2013 the official showcase for the Gemma open-model ecosystem. The feature validates Hirundo<\/a>\u2019s weight-level machine unlearning approach as a production-grade solution to one of the most persistent vulnerabilities in enterprise AI deployments: prompt injection attacks.<\/p>\n The hardened model is built on Google\u2019s Gemma 4 E4B instruction-tuned base. At 4 billion parameters, it outperforms models more than 170 times its size on prompt injection resistance \u2013 including DeepSeek V3.2-Exp (685B), GPT-OSS-120B, and Qwen3-235B \u2013 while preserving full utility across standard benchmarks.<\/p>\n AI security is a behavioral problem, not a size problem<\/p>\n Prompt injection \u2013 where adversarial inputs manipulate a model into overriding its system instructions \u2013 remains one of the most damaging attack vectors in production Large Language Model (LLM) deployments. The prevailing assumption has been that larger models are inherently more robust. Hirundo\u2019s results directly contradict that.<\/p>\n Rather than applying external filters or inference-time guardrails, Hirundo\u2019s platform targets and removes the specific model weights responsible for susceptibility to adversarial manipulation. The model effectively forgets the behaviors that enable attacks \u2013 at the weight level, not the prompt level. Despite those behaviors usually being tied to instruction-following, a desired trait of LLMs, Hirundo\u2019s technology preserved general instruction following while mitigating prompt injection susceptibility.<\/p>\n The results against industry-leading open-weight models, benchmarked using Meta\u2019s PurpleLlama CyberSecEval dataset:<\/p>\n DeepSeek V3.2-Exp (685B): 73.33% attack success rate \u2013 15.6x worse than Hirundo\u2019s model<\/p>\n GPT-OSS-120B: more than 3x Hirundo\u2019s attack success rate<\/p>\n Qwen3-235B: 10.8x more vulnerable<\/p>\n Hirundo\u2019s hardened Gemma 4 E4B achieved a 4.78% attack success rate \u2013 a 74.47% reduction versus the base model \u2013 with no measurable degradation across utility benchmarks including AIME25, LiveCodeBench, GPQA, IFBench, SCICode, AutoPatchBench, and CyberSOCEval.<\/p>\n Quote<\/p>\n “Prompt injection is not a prompting problem \u2013 it is a representational one. The vulnerability lives in the weights. Addressing it at the weight level is more durable and more precise than guardrails applied after the fact.”<\/p>\n \u2013 Prof. Em. Oded Shmueli, Chief Scientist, Hirundo; former Executive VP for Research and Dean of the Computer Science Faculty, Technion \u2013 Israel Institute of Technology<\/p>\n The Google DeepMind-endorsed Gemma 4 case study joins a growing portfolio of Hirundo\u2019s publicly available debiased and security-hardened models, including variants of GPT-OSS, Qwen, NemoTron, Llama, and others \u2013 with results demonstrating up to 90%+ improvement in security robustness and fairness compared to the original base models.<\/p>\n Availability<\/p>\n Hirundo\u2019s Gemma 4 E4B Unlearned Variant on HuggingFace: Hirundo\u2019s Gemma 4 E4B Case Study on Google DeepMind\u2019s Gemmaverse: About Hirundo<\/p>\n Hirundo is the pioneer of machine unlearning \u2013 making it possible to change what trained AI models know and how they behave. Rather than applying filters or guardrails after the fact, Hirundo\u2019s platform operates at the weight level, enabling teams to remove sensitive data, harmful behaviors, jailbreak vulnerabilities, and bias at the core \u2013 with no measurable degradation to model utility.<\/p>\n The company holds nine filed US patents and is led by Ben Luria (CEO), one of the first Israeli Rhodes Scholars; Prof. Em. Oded Shmueli (Chief Scientist), former Executive VP for Research and Dean of the Computer Science Faculty at the Technion \u2013 Israel Institute of Technology; and Michael Leybovich, an award-winning researcher.<\/p>\n www.hirundo.io<\/a><\/p>\n View source version on businesswire.com: https:\/\/www.businesswire.com\/news\/home\/20260521997520\/en\/<\/a><\/p>\n Contacts<\/p>\n
https:\/\/huggingface.co\/hirundo-io\/gemma-4-E4B-it-reduced-prompt-injection<\/a><\/p>\n
https:\/\/deepmind.google\/models\/gemma\/gemmaverse\/hirundo<\/a><\/p>\n