Ismail Kharoub

CTO and Co-Founder of VlnoAI

[Ismail Kharoub @IsmailOfVlno]

@gilad2323 אין פרמטרט אחד שג׳ימני מוביל בו

[Ismail Kharoub @IsmailOfVlno]

@htokgozk @trq212 We did well there didn’t we 😉

[Hertzel Kuriel @htokgozk]

Today I'm proud to share something we've been building toward, alongside a partner we're proud to build with:Wand AI

Wand AI just announced that vlno ai is now the model robustness layer in their Sovereign AI stack, so governments and regulated enterprises can run open-weight models on their own infrastructure with robustness they can actually measure, not just assume.

Here's the problem we exist to solve. Every organization wants the control and economics of open models. But an agent built on a model can be manipulated by the content it reads into misusing the access it was given, and that risk is what stops these deployments in security review. Runtime filters and red-team reports can find the weakness. Neither changes what the model does next.

That's our whole reason for being. VLNO tests a model against adversarial adaptive attacks and produces the training data that hardens the model itself, so the robustness lives in the weights and travels with the model wherever it runs. Your existing controls stay in place on top. As the model changes, we re-test and re-harden, so the protection stays current.

In practice, a ministry or a bank can put an open model into production on a measured attack success rate, not a vendor's promise. Robustness becomes something a model has to earn before it reaches production, and keeps earning over time.

To my co-founder @IsmailOfVlno and our team: you turned a hard idea into something real. Big thanks to the Wand AI team for starting this with us. We're just getting started.

And to the enterprises and teams putting agents into production: we'd love to show you what measured robustness looks like for your models.

Link to PR in the first comment.

We train the threat out.

[Ismail Kharoub @IsmailOfVlno]

@daardos K3 לא נופל ממנו

[Ismail Kharoub @IsmailOfVlno]

@urieli17 לא צריך טרליון פרמטרים כדי לבנות מודל מומחה, והם בוודאות לא התחילו מאפס

[Ismail Kharoub @IsmailOfVlno]

@harleyfoote_ @0x0SojalSec Those specifically were opus 4.7 designed injections from a held out benchmark, not part of the training data

[Thariq @trq212]

we should have called this post "defeating the lethal trifecta"

https://t.co/iGPfHR9I9y

[Ismail Kharoub @IsmailOfVlno]

@mbrg0 This is cool, but Haiku 4.5 is not sonnet 5

[Ismail Kharoub @IsmailOfVlno]

@urieli17 השתתפי במחקר שהוביל לזה, יאמר לזכותם שהם עשו אחלה עבודה הפעם

[Ismail Kharoub @IsmailOfVlno]

@buzagloidan @alono88 אפשר לשכנע דפדפן להדליף את המידע שלך החוצה?

[Ismail Kharoub @IsmailOfVlno]

@buzagloidan @alono88 ככה

[Ismail Kharoub @IsmailOfVlno]

@Lihi4Lihi מתחת לרדר?

[Ismail Kharoub @IsmailOfVlno]

@DeanMeyerrr Great read, AI security needs to evolve from static audits to continuous behavioral verification.

I think supply-chain trust won’t come from open weights alone, it’ll come from continuously proving that a model still satisfies its expected security and safety properties after PT

[Ismail Kharoub @IsmailOfVlno]

@alono88 שלא נדבר על זה שהוא א��ד הלא מאובטחים

[Ismail Kharoub @IsmailOfVlno]

@NateBubis @alono88 ��Don’t make security mistakes”

[Ismail Kharoub @IsmailOfVlno]

@VittoStack Sounds very interesting, happy to join

[Ismail Kharoub @IsmailOfVlno]

@shaulig אתה מרגש יא זבלון, אני שמח שאתה פה ❤️

[Ismail Kharoub @IsmailOfVlno]

@shaulig שם מקסים ❤️

[Ismail Kharoub @IsmailOfVlno]

יססס!!!

[Shaul Greenfeld @shaulig]

אני בן שמונה חודשים. לומד הכל מחדש.
בעיקר להתחיל ללכת.
אני הולך.