--- title: Don't Trust the Salt: AI Summarization, Multilingual... description: A researcher at Taraaz presents three interconnected projects exposing critical weaknesses in LLM safety systems across non-English languages. The 'Bilingual... canonical: https://daily.dev/posts/don-t-trust-the-salt-ai-summarization-multilingual-safety-and-the-llm-guardrails-that-need-guardi-zyr7yr1yw twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Don't Trust the Salt: AI Summarization, Multilingual Safety, and the LLM Guardrails That Need Guarding | daily.dev og:description: A researcher at Taraaz presents three interconnected projects exposing critical weaknesses in LLM safety systems across non-English languages. The 'Bilingual... og:url: https://daily.dev/posts/don-t-trust-the-salt-ai-summarization-multilingual-safety-and-the-llm-guardrails-that-need-guardi-zyr7yr1yw og:image: https://api.daily.dev/og/posts/ZyR7Yr1YW.png og:image:alt: Don't Trust the Salt: AI Summarization, Multilingual Safety, and the LLM Guardrails That Need Guarding og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Don't Trust the Salt: AI Summarization, Multilingual Safety, and the LLM Guardrails That Need Guarding **[Hacker News](https://daily.dev/sources/hn)** · 10 min read · 1 upvotes · 0 comments ## Summary A researcher at Taraaz presents three interconnected projects exposing critical weaknesses in LLM safety systems across non-English languages. The 'Bilingual Shadow Reasoning' technique demonstrates how customized non-English system prompts can steer a model's hidden chain-of-thought to bypass safety guardrails while producing seemingly neutral outputs — tested on GPT-OSS-20B using a Farsi policy mirroring authoritarian framing of human rights. A Multilingual AI Safety Evaluation Lab (built at Mozilla Foundation) benchmarked GPT-4o, Gemini 2.5 Flash, and Mistral Small on refugee/asylum scenarios in Arabic, Farsi, Pashto, and Kurdish, finding significant quality drops versus English, inconsistent safety disclaimers, and LLM-as-a-Judge inflating scores and hallucinating disclaimers. A third project tested whether guardrail tools (FlowJudge, Glider, AnyLLM) actually enforce multilingual policies, finding Glider produced 36–53% score discrepancies based solely on policy language. The core argument: evaluation insights must flow directly into guardrail design, and the safety tools themselves are failing in non-English contexts. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Similar posts on daily.dev - [19 large language models for safety or danger](https://daily.dev/posts/19-large-language-models-for-safety-or-danger-ki6rjj3pu) · InfoWorld · 1 upvotes · 0 comments - [An LLM that will help you build a nuclear weapon](https://daily.dev/posts/an-llm-that-will-help-you-build-a-nuclear-weapon-r1yxtwj5e) · InfoWorld · 0 upvotes · 0 comments - [Researchers discover a shortcoming that makes LLMs less reliable](https://daily.dev/posts/researchers-discover-a-shortcoming-that-makes-llms-less-reliable-wiiiapmcn) · MIT News · 2 upvotes · 0 comments - [NDSS 2025 – Safety Misalignment Against Large Language Models](https://daily.dev/posts/ndss-2025-safety-misalignment-against-large-language-models-hpoay68cq) · Security Boulevard · 0 upvotes · 0 comments --- Tags: [#red-teaming](https://daily.dev/tags/red-teaming) [View this post on daily.dev](https://daily.dev/posts/don-t-trust-the-salt-ai-summarization-multilingual-safety-and-the-llm-guardrails-that-need-guardi-zyr7yr1yw) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Don't Trust the Salt: AI Summarization, Multilingual Safety, and the LLM Guardrails That Need Guarding","url":"https://daily.dev/posts/don-t-trust-the-salt-ai-summarization-multilingual-safety-and-the-llm-guardrails-that-need-guardi-zyr7yr1yw","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/don-t-trust-the-salt-ai-summarization-multilingual-safety-and-the-llm-guardrails-that-need-guardi-zyr7yr1yw"},"datePublished":"2026-02-19T14:12:02.244Z","dateModified":"2026-02-19T14:12:28.329Z","description":"A researcher at Taraaz presents three interconnected projects exposing critical weaknesses in LLM safety systems across non-English languages. The 'Bilingual...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/1aabe0a34ca945426acda5d935202656?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/1aabe0a34ca945426acda5d935202656?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/don-t-trust-the-salt-ai-summarization-multilingual-safety-and-the-llm-guardrails-that-need-guardi-zyr7yr1yw","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"red-teaming","timeRequired":"PT10M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Don't Trust the Salt: AI Summarization, Multilingual Safety, and the LLM Guardrails That Need Guarding"}]} ```