Bynario's €2.1M: AI found Apple's bug first

Yesterday we wrote about Accomplish's Beltdown disclosure and the question its CTO asked out loud: if these models are so good, why aren't they finding critical bugs in their own vendors' products? Today's answer comes from Milan. Bynario pointed frontier models at Apple's operating systems, found real flaws in macOS Screen Sharing, and turned the result into a €2.1 million pre-seed round. AI vulnerability discovery stopped being a conference demo this week and became a fundable product.

What actually happened

Bynario closed a €2.1 million pre-seed led by 360 Capital, with PranaVentures participating, according to EU-Startups and Tech.eu. The company was founded in 2025 by CEO Pesoli alongside Giancarlo Russo and Lorenzo Cavallaro.

The proof point came first and the round came after. Bynario's researchers used frontier AI models to uncover security flaws in Apple's macOS Screen Sharing technology. The CEO estimated the underground market value of the resulting exploit chain at somewhere between $100,000 and $200,000 — his figure, not an audited number, but it tells you the class of bug they were working at.

The product is a continuous exposure management layer: identify, validate, prioritize, and remediate vulnerabilities from code through cloud. The new money goes to the platform, engineering headcount, and enterprise scaling.

Note the four verbs in that list. Only one of them is discovery.

Why this matters for your business

The discovery half of this is already commoditized, and that is the part that should concern you. Anyone can point a good model at a public repository. That includes whoever is looking at yours. The economics of finding a bug in your code dropped this year, and your patch cadence did not speed up to match.

Validation is the whole product. Turn a frontier model loose on a mid-sized codebase and it returns hundreds of findings, most of them wrong or unreachable. An AI scanner without a validation step is a queue-filling machine, and a queue nobody trusts gets ignored — which is worse than no scanner. When a vendor pitches you AI appsec, ask what fraction of findings they prove exploitable, and ask to see a false-positive rate.

Prioritization beats volume. Your real question is not "how many issues do we have," it is "which two do we fix this week." Any tool that cannot rank against your actual deployment — what's internet-facing, what holds credentials, what runs in CI — is handing you homework instead of an answer.

Assume the asymmetry. Attackers running the same models do not need validation. A false positive costs them one wasted afternoon. It costs your team a sprint. Close the gap by shrinking what an exploitable bug can reach: scoped credentials, short-lived tokens, and secrets that do not live on the box that serves your code.

Sources: EU-Startups, Tech.eu.