Multilingual phishing detection still fails where language cues matter

At a glance

What this is: NHI Mgmt Group's analysis, drawing on Abnormal AI's blog post: Multilingual phishing detection fails when models depend on translation rather than cultural and linguistic nuance.

Why it matters: IAM and security teams need detection and training controls that understand local language patterns, or localized BEC and phishing will continue to bypass English-first guardrails.

Context

Multilingual phishing becomes a governance problem when detection systems are built around English syntax and translation rather than how people actually write in each market. That gap matters because attackers use language-specific cues, including formality, tone, and honorifics, to make fraudulent messages appear legitimate.

Abnormal AI argues that this is not just a content-classification issue but a model-design issue: the control has to understand business communication in context, across languages and cultures. For security programmes operating globally, the question is whether email security and awareness tooling can still distinguish malicious intent when the attack is linguistically local.

Key questions

Q: How should security teams test phishing detection across different languages?

A: They should measure detection and false positive rates by language, not just globally, and validate against real business communication in each region. The test set should include native phrasing, honorifics, tone shifts, and impersonation patterns that reflect how employees and suppliers actually write.

Q: Why do English-first phishing controls miss localized BEC attacks?

A: Because they often score messages on translated wording and link patterns while missing the social meaning of local communication. Attackers exploit formality, politeness, and relationship cues that are obvious to native readers but invisible to systems trained mainly on English.

Q: What are the signs that multilingual email security is underperforming?

A: Look for higher false negatives, inconsistent alert quality, and more manual review in non-English traffic than in English traffic. Another warning sign is when users in regional teams report obvious scams that the SOC never saw as suspicious.

Q: How can organisations align awareness training with multilingual detection?

A: Use native-language simulations that mirror the same cues the detection engine is meant to catch, including tone, urgency, and impersonation style. Training should not rely on generic translated templates if the production control is expected to govern local-language attacks.

Technical breakdown

Why translation-based phishing filters miss local language cues

Translation-based detection normalizes text into a form that strips away the very cues attackers exploit. In German, suspiciously informal phrasing can signal impersonation or social engineering, while in Japanese the wrong level of keigo, or honorific speech, can make a message read unnatural to a native speaker even if the translated content looks harmless. English-trained systems often overfit to vocabulary and link patterns, so they miss tone, register, and relationship cues that matter in real business communication. That creates a false sense of coverage when the attack is culturally precise.

Practical implication: tune detection to multilingual communication features, not just translated keywords.

How multilingual embeddings change detection logic

Abnormal describes a shift from rule-based text tagging to a neural classifier using multilingual embeddings. In practical terms, embeddings allow a model to represent language in a way that captures contextual similarity across words, phrases, and communication style rather than relying on literal matches. That matters for phishing because malicious intent often hides in formality mismatches, impersonation patterns, and odd relationship cues instead of obvious payload indicators. When the system is trained on authentic regional data, it can compare legitimate and malicious examples in the language as it is actually used, not as it appears after translation.

Practical implication: validate whether your detection stack is trained on regional communication data, not synthetic translations.

Why native-language training needs to match native-language detection

Detection and awareness training fail in different ways when they are both built around English assumptions. If the inbox control cannot identify a localized phish, users may never see a relevant example, and if the training content is templated or translated without nuance, employees learn the wrong cues. Native-language simulations close that loop by teaching people to spot the same tone, urgency, and impersonation tactics the detection engine is trying to catch. That turns localization from a presentation layer into an operational control across both prevention and response.

Practical implication: align simulation content with the languages and communication styles your detection engine is expected to govern.

Threat narrative

Attacker objective: The attacker wants to persuade the recipient that the message is locally authentic so the fraud can proceed without triggering standard email controls.

1. Entry occurs through a localized phishing or BEC message written to sound native to the target, using language-specific cues to evade English-first filters.
2. Credential or engagement harvesting follows when the recipient is steered toward a reply, callback, or lookalike login page that appears culturally credible.
3. Impact is the successful impersonation of trusted people or services, which increases the chance of fraud, account compromise, or off-channel manipulation.

Breaches seen in the wild

• CoPhish OAuth phishing via Copilot Studio: Datadog showed Copilot Studio agents on a Microsoft domain can front OAuth consent phishing and forward stolen tokens; no victims reported.
• Twilio 0ktapus breach 2022: SMS phishing of employees exposed 209 Twilio customers and 1,900 Signal users, part of the 0ktapus campaign against 130+ firms.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 150+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

Localized phishing is an identity governance problem, not just a content problem. Security teams often assume that translation coverage equals detection coverage, but this article shows that language formality, honorifics, and tone are part of the attack surface. A control that cannot interpret those signals cannot reliably distinguish legitimate communication from impersonation, which means multilingual coverage has to be treated as a governance requirement, not a localisation feature.

Communication nuance is a named control gap: multilingual intent blindness. The model failure is not that systems cannot read foreign text, but that they cannot read the social meaning carried by local phrasing. That matters across BEC, phishing, and executive impersonation because attackers use language as an access path to trust. Practitioners should read this as evidence that detection quality in global environments depends on region-specific training data and model behaviour, not translation alone.

Localized detection and localized awareness need to evolve together. If the inbox layer learns multilingual nuance while user training remains generic, the programme still leaves a gap between what the SOC sees and what employees are taught to notice. The article supports a broader point: global identity security programmes must govern communication patterns end to end, from detection to simulation, or the weakest language becomes the easiest entry point.

English-first assumptions distort global assurance. Many organisations treat multilingual support as a coverage checkbox, but the operational question is whether controls behave consistently in the markets where the business actually operates. The evidence here suggests the answer is often no, which means security leaders should evaluate language-localised detection and training as part of their assurance baseline, not as an optional enhancement.

What this signals

Multilingual phishing is exposing a programme design flaw: security teams still treat language as presentation, when it is part of the attack itself. Detection, simulation, and analyst review all need to understand local business communication if global coverage is meant to be real.

Native-language control alignment is becoming the practical test for global email security. If a programme cannot measure and respond to German, Japanese, and other local forms of impersonation with comparable confidence, it is not globally governed, only globally deployed.

For practitioners

• Audit language coverage in email security controls Test whether detection quality drops in the languages your organisation actually uses for suppliers, finance, executive communication, and internal approvals.
• Replace translation-only phishing models Require models that evaluate tone, register, and impersonation cues in the source language, especially for German and Japanese business traffic.
• Use native-language simulations in security awareness Build training examples in the languages employees receive in daily work so the lessons reflect local phrasing, urgency, and honorific mistakes.
• Benchmark false negatives by language Track missed threats and false positives separately for each major language to identify where tuning or training data still lags.
• Treat localization as a control requirement Make multilingual performance part of your acceptance criteria for email protection and awareness programmes rather than a post-deployment enhancement.

Key takeaways

• Localized phishing succeeds when security tools treat translation as a substitute for understanding how people actually communicate in each language.
• The article shows that tone, honorifics, and formality are not cosmetic details. They are operational signals attackers use to make BEC and phishing messages appear legitimate.
• Global programmes should align detection, awareness, and validation by language so regional gaps are visible before attackers use them.

Key terms

• Multilingual phishing: Phishing that is written to match the language, tone, and business style of the target audience. The attacker relies on local cues such as formality, honorifics, and phrasing to make malicious messages look routine, which makes simple translation-based detection less effective.
• Honorific Speech: Honorific speech is language that signals respect, hierarchy, and social distance, especially in languages such as Japanese. In security operations, the wrong level of honorifics can be a useful indicator of impersonation or synthetic writing, because attackers often miss the social rules native speakers use naturally.
• Business email compromise: A form of social engineering where an attacker impersonates a trusted person or domain to manipulate payment, change banking details, or extract sensitive information. It often succeeds without malware because the attacker targets process trust and human judgement instead of technical controls.
• Multilingual embeddings: A machine learning representation that maps words and phrases from different languages into a shared semantic space. In security, it helps models compare meaning across languages instead of depending on exact wording, which improves detection of nuanced social engineering.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity lifecycle are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are building or maturing an IAM programme, it is worth exploring.