Hirundo is developing industry-first solutions to “make AI forget” poisoned, malicious and confidential data from trained AI models, biases as well as hallucinations, the problem of AI models generating misleading or inaccurate information that appears factual. Hirundo achieves this without needing to retrain models from scratch and without degrading model performance, thereby de-risking AI adoption.
The company has raised an $8M seed funding round, led by Maverick Ventures Israel, with participation from SuperSeed, Alpha Intelligence Capital, Tachles VC, AI.FUND and Plug and Play Tech Center.
As generative AI evolves from experimentation to enterprise infrastructure, critical flaws start to appear, and once a model is trained, fine-tuned or deployed, its mistakes are nearly impossible to fix. Researchers have found almost 40% of ‘facts’ used by AI contain bias, while half of US employees cite inaccuracy as a major concern associated with generative AI. As AI becomes more pervasive and more powerful, the more hallucinations and biases are uncovered, with hallucinations worse now more than ever before, leading to legal, reputational and operational risks for businesses relying on it. Hirundo is addressing the flaws in AI from an entirely unique starting position – after a model is already trained. By retroactively removing undesired behaviors and inaccuracies from AI models without impeding the model’s capabilities or needing to retrain it – a time-consuming and multi-million dollar exercise – Hirundo improves AI performance and reliability, empowering businesses to deploy AI with confidence.
Techniques to remediate issues in AI models are varied and span the entire AI development lifecycle, but these still do not prevent hallucinations or biases from leaking into final AI projects. Conventional solutions like guardrails or fine-tuning are insufficient – they simply mask and filter bad data or behaviors while the root cause in the model itself remains untouched. Hirundo looks at jailbreaks from a behavioral point of view – using its machine unlearning approach, the company locates the directions in the model that are most prone to adversarial manipulations, and removes them. This aims to future-proof AI models to attacks.
“Broader adoption of AI is limited by hallucinations and undesired behaviors which make models too risky to deploy in enterprise-level applications. With Hirundo, models can be remediated instantly at their core, working towards fairer and more accurate outputs,” said Ben Luria CEO & Co-Founder of Hirundo. “Hirundo’s solution operates like a form of AI model “neurosurgery,” pinpointing where in a model’s billions of parameters hallucinations originate or toxic knowledge encoded, and precisely removing it. We ensure data is reliably deleted, model accuracy is assured, and the process is scalable and repeatable. We welcome this latest funding which will enable us to enhance our technology and with it, support the broader adoption of AI.”
“Without removing hallucinations or biased intelligence from AI, we end up distorting outcomes and encouraging mistrust – for organizations and enterprises this means legal, financial and real-world consequences. A majority of businesses suffer from AI built with bias, but most cannot address it sufficiently. Hirundo offers a type of AI triage; removing untruths or data built on discriminatory sources and completely transforming the possibilities of AI. We are delighted to invest and support their onward journey,” said Yaron Carni, Founder at Maverick Ventures Israel, which led this funding round.
Mads Jensen, Managing Partner of SuperSeed commented; “We invest in exceptional AI companies transforming industry verticals, but this transformation is only as powerful as the models themselves are trustworthy. Hirundo’s technology is groundbreaking in addressing the hallucination and bias problems that plague even the most advanced models today. Their approach to machine unlearning addresses a critical gap in the AI development lifecycle.”
Hirundo identifies how a particular behavior presents in the model, utilizing industry accepted benchmarks and outlines any undesirable behavioral traits, followed by steering the model away from them. That approach works both for open-source models (like Llama, Mistral, Gemma, etc.) and soon, gated models (like ChatGPT, Claude, etc.). When deployed, Hirundo’s solution has led to the removal of up to 70% of biases – as demonstrated in their work on DeepSeek-R1, as well as up to 55% reduction of hallucinations and 85% decrease in successful prompt injections, showcased with the company’s work on Llama.
The first-choice solution for mission-critical AI deployments in industries such as finance, healthcare, and other consumer-facing industries, as well as high-risk enterprise and defense applications, Hirundo is already piloting with a variety of multinational corporations and government agencies. Hirundo’s platform supports both generative models (like large language models) and non-generative systems used in areas such as computer vision, radar, LiDAR, NLP, and more.
Founded in 2023 by Emeritus Professor Oded Shmueli, former Dean of Computer Science and EVP at Israel’s Technion, alongside Ben Luria, serial entrepreneur and Rhodes Scholar, and Michael Leybovich, an expert in the field of data lineage, Hirundo is the first startup in the world to offer an unlearning solution, pioneering the concept of “making AI forget.”
