Cybersecurity’s next front line is multilingual

AI engineers are developing detection systems that look beyond English to identify threats and sensitive information

A phishing email written in fluent Japanese lands in the inbox of a finance team in Tokyo. A contractor in São Paulo pastes a confidential pricing sheet into a chatbot and asks for a summary in Portuguese. Neither scenario involves a word of English, raising a question for businesses using AI security tools: How well do those protections work when the language changes?

The stakes are growing. Gartner forecast worldwide information security spending would reach $240 billion in 2026, up 12.5% from a year earlier. The research firm identified expanding AI use among businesses and attackers as a growth driver.

Kaustav Saha, a senior staff artificial intelligence engineer and architect at Zscaler, has been working on that language challenge. His work focuses on detection systems designed to handle 26 languages and their scripts, including efforts to identify sensitive information even when it has been paraphrased or translated.

Saha says the effort has produced five AI white papers and five U.S. patent filings within six months. Two of the designs are publicly available as defensive publications, which describe technical approaches rather than independently establish their effectiveness.

Threats do not wait for translation

Security systems built primarily around English can encounter problems when extended to other languages. Translating content before analyzing it may change the cues a detector needs, such as a company name transliterated into Devanagari or an account number formatted to local conventions.

Attackers and employees write in the languages they use every day, and a multinational company’s data moves through many of them.

Saha says his team treats each language as a direct input rather than relying on translation into English first. His work includes multilingual detection, with plans to extend that approach to images and conversations spanning multiple messages.

One published design combines a multilingual transformer encoder, which creates a numerical representation of text, with features that help identify names and organizations. A classifier then evaluates whether a message mentions a particular entity.

A script-aware filter adjusts parts of the process for non-Latin scripts, where familiar English-language cues such as capitalization may not apply.

“I grew up reading Bengali and Hindi alongside English, so I never thought of the rest of the world as an edge case,” Saha said. “If you translate first and detect second, you have already thrown away half the evidence.”

Catching a leak that shares no words with the source

Employees can expose sensitive information when they paste contracts, pricing documents or source code into an AI tool. Detecting that exposure becomes more complicated when the text has been rewritten or summarized in another language.

Exact-match detection looks for text that matches protected material. A system designed to recognize meaning must address a different problem: whether a prompt carries sensitive information without repeating the original wording.

A technical disclosure co-authored by Saha and colleagues describes a system that evaluates three signals: similarity in meaning, overlap in named entities, and matching numbers or identifiers.

The design combines those signals using weights selected for the detected language. Clear cases follow faster decision paths, while ambiguous cases receive an additional comparison through a more computationally intensive model.

The resulting assessment indicates whether the input reproduces information from the protected reference material in meaning. The disclosure describes the architecture, not an independently verified detection rate.

“A leak rarely looks like a copy,” Saha said. “Someone rewrites a paragraph in German and changes the currency format, and a string matcher sees nothing. We built the system to ask a different question: does this text carry the same facts?”

Holding languages to the same standard

Coverage claims are easy to make and harder to evaluate. An overall accuracy figure can conceal differences between languages, especially when some have far less training data than others.

False positives also matter. A filter that repeatedly blocks legitimate business email can disrupt work and undermine confidence in the protection.

Saha argues that production systems need to be measured language by language, including how quickly they make decisions under live traffic conditions.

He says his multi-topic detection work uses the mmBERT multilingual encoder and combines two detectors under a shared classification framework to evaluate content across 26 languages.

That emphasis on practical testing also informs his assessment of other engineers’ work. Saha said he served as a judge at the Global Hackathon for Data & AI Engineers, held Sept. 15-17, 2026.

“When everyone in the room has run production systems, a demo that only works on the happy path does not survive the first question,” Saha said. “I hold my own models to the same test. If a language has not been measured, I do not count it as covered.”

The same exposure reaches global products

The challenge extends beyond security vendors. Companies putting chatbots or AI agents in front of international customers need to consider how those systems respond across languages.

A brand name may also be an ordinary word, a place name or the name of an unrelated business. Recognizing the name alone does not establish whether a message presents a threat.

Saha describes his entity-anchored multilingual brand-risk detection work, known as EAMBR, as an attempt to separate those cases. The proposed approach connects an alert to a specific entity, distinguishes similar names and evaluates risk using categories and signals of apparent intent.

Those are design goals, rather than proof that a system will reliably interpret every ambiguous message.

“The teams I respect most start with the failure cases,” Saha said. “A brand name that is also an ordinary word in Spanish, or a product that shares its name with a town in Indonesia. Solve those and the easy cases take care of themselves.”

The conversation becomes the attack surface

The scale of non-English threats is already visible. In a May 2025 analysis, Proofpoint researchers documented CoGUI phishing campaigns primarily targeting users in Japan. The researchers observed more than 172 million messages associated with the kit in January 2025 alone, with campaigns impersonating consumer and financial brands.

Saha expects detection challenges to extend beyond individual messages to requests spread across several conversational turns, potentially combining text and images.

He describes his latest design, DISPATCH, as a dual-stream architecture that evaluates intent and semantic content in parallel before applying policy-based thresholds.

His broader work also includes a book on engineering large language models for non-Latin scripts. The common focus is building systems that handle those scripts directly, rather than treating English as the necessary intermediate step.

For businesses evaluating these tools, the practical questions extend beyond how many languages appear on a coverage list. They include how each language was tested, how often legitimate work is blocked and whether the system can recognize sensitive information after its wording changes.