Hirundo announced an $8 million seed round on June 9, 2025, to advance software designed to remove or reduce unwanted behavior in trained AI models. The round was led by Maverick Ventures Israel, with SuperSeed, Alpha Intelligence Capital, Tachles VC, AI.FUND and Plug and Play Tech Center also participating. Hirundo describes its approach as machine unlearning: modifying a model after training rather than retraining it from scratch.
What Hirundo says its software does
Founded in 2023 by Ben Luria, Michael Leybovich and Oded Shmueli, Hirundo develops enterprise software intended to identify and change unwanted information or behavior learned by an AI model. The company says its system targets issues such as hallucinations, bias, jailbreaks and prompt injections, toxic outputs, and memorized personal or confidential information.
Hirundo positions the product for use before a model launches, when problems emerge in production, and as part of ongoing model hardening. Its site offers a demo and an early-access sign-up; the reviewed company materials do not state public pricing or describe a self-serve purchase option. Hirundo’s product information is the source for these availability and product descriptions.
How machine unlearning differs from filters or retraining
In Hirundo’s framing, output filters and guardrails act around a model by controlling what it returns, while its unlearning method is meant to modify the trained model itself. The company also contrasts its approach with retraining from scratch, which it characterizes as resource-intensive. Those are Hirundo’s product comparisons, not independent findings that filters or retraining are ineffective.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
Changing a model to suppress a behavior also raises a practical question: does the intervention reduce the targeted risk without damaging other capabilities? The announcement does not establish that utility is preserved in every context. Organizations evaluating any model-unlearning product should ask which model families and deployments are supported, how the change is measured, and whether performance on unrelated tasks is tested.
What Hirundo reports about results
In its June 9, 2025 announcement, Hirundo reported the following results. The figures are company claims, not general guarantees:
| Reported outcome | Model identified by Hirundo | Qualification |
|---|---|---|
| Up to 55% fewer hallucinations | Llama | Reported by Hirundo in 2025; the announcement does not provide independent replication or detailed benchmark protocols. |
| Up to 70% reduction in bias | DeepSeek-R1 | Reported by Hirundo in 2025; the announcement does not provide independent replication or detailed benchmark protocols. |
| 85% decrease in successful prompt injections | Llama | Reported by Hirundo in 2025; the announcement does not provide independent replication or detailed benchmark protocols. |
The named models and company attribution matter: these percentages do not show that the same reductions apply to arbitrary models, benchmarks or production environments. The reviewed sources do not provide enough detail to independently assess the test conditions or determine whether other model capabilities changed.
What the funding announcement establishes—and what it does not
The financing is a seed round for a business software company, not a consumer product launch. Hirundo’s announcement identifies Maverick Ventures Israel as lead investor and names SuperSeed, Alpha Intelligence Capital, Tachles VC, AI.FUND and Plug and Play Tech Center as participants. Funding can support a company’s development, but the round itself does not validate product efficacy or prove commercial traction.
Recommended Free Tools
Rank #3
Hirundo CEO and co-founder Ben Luria described the approach as “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.” That is the executive’s analogy and claim, not independent technical validation. Likewise, the investor commentary in the announcement reflects an investment rationale rather than an independent evaluation.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What enterprise buyers should verify
For a company considering machine unlearning, the useful comparison is not just a headline percentage. Ask vendors for evidence that fits the intended model and deployment:
- Intervention: Does the method change model parameters, filter outputs externally, or combine both?
- Target and scope: Which trained model, behavior and information are addressed, and does the method require retraining?
- Evaluation: Which benchmarks and test conditions support the claimed result, and are results independently replicated?
- Utility: What happens to accuracy and behavior on tasks unrelated to the targeted issue?
- Deployment and cost: Which model families and deployment settings are supported, and what are the costs, latency and production requirements?
The available sources are not sufficient for ranking Hirundo against other providers on these dimensions. Organizations can contact Hirundo about a demo or early access, then request evidence tied to their own models and risk scenarios.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errors
