Habr•November 17, 2026•🇷🇺Translated from Russian

Tokenization Exploits: How One Unicode Comma Forces LLMs to Switch Languages and Bypass Filters

Tokenization and embedding layers in large language models create attack surfaces that exist entirely before the residual stream begins. Researchers tracing a single token from raw Unicode string through BPE encoding have shown that one three-byte punctuation mark can hijack an entire generation trajectory.

Sending the prompt hello, deepseek consistently produces English replies. Replacing the comma with the full-width variant hello,deepseek causes the model to respond in Chinese in 94 percent of cases, even though the first token inside the thinking block remains Hello. The semantic content is identical; only the byte sequence reaching the embedding matrix changes.

The tiktoken library used by most frontier models applies a fixed four-stage pipeline: pre-tokenization via complex regex, UTF-8 byte conversion, learned BPE merges, and final ID lookup. Because the merge ranks are frozen after training, any input that crosses pre-tokenization boundaries produces different token IDs than defenders expect.

Controlled API experiments with DeepSeek (deepseek-flash, temperature 0.7) confirmed the effect across 400 randomized queries. The full-width comma alone drove Chinese output in 95 of 100 trials without a following space and 93 of 100 trials with a space. Standard ASCII commas produced zero Chinese responses.

Similar behavior appears in Claude and ChatGPT web interfaces when the same punctuation substitution is applied. The models themselves attribute the language switch directly to the comma character.

Additional vectors include Zalgo-style diacritic stacking that expands a 64-character prompt into 189 tokens and homoglyph substitution that fragments known words such as amazon.com into multiple low-frequency tokens. These techniques enable both context-shadowing against upstream classifiers and KV-cache exhaustion attacks.

The findings demonstrate that any security mechanism relying on Unicode categories, character counts, or surface-form inspection will diverge from the actual token sequence consumed by the model, creating reliable bypass opportunities at the embedding layer.

Related articles

AntiMalware•AI Security

Housefishing Emerges as Sellers Use Generative AI to Mislead Buyers on Rightmove

A new form of real estate deception called housefishing is spreading across the United Kingdom, where sellers edit property photographs with neural networks or replace them entirely with AI-generated images. The practice draws its name from catfishing and has been documented by BBC reporting on the experiences of homebuyer Olivia Nicole. Nicole discovered significant discrepancies during an in-person viewing on the Rightmove platform, including smaller rooms, altered window sizes, and a wall protrusion hidden behind a sofa. Real estate blogger Jack Bushrod has observed a sharp rise in such manipulated listings over the past six months, with most failing to disclose the AI alterations. Common edits include removal of structural columns, addition of nonexistent windows, and concealment of carpet stains or tile cracks. Industry experts consider limited AI use acceptable for removing clutter or staging empty rooms, provided the changes are clearly labeled.

Habr•AI Security

AI Framework ARTEX Actively Used in Real-World Attacks for Cloud Credential Theft and Crypto KYC Bypass

Multiple sources including Amazon and the FBI reported active exploitation of the open-source AI framework ARTEX in real cyberattacks. Written in Go, ARTEX leverages external LLMs such as Claude and ChatGPT to perform reconnaissance, vulnerability exploitation, and data theft. Originally published on GitHub in July 2026 as a penetration testing tool, the repository was removed by its author in October after reports of malicious use. Amazon researchers detected large-scale campaigns where ARTEX was used to steal access keys to AI agents through specially crafted AI honeypots. A second campaign targeted Know Your Customer checks in cryptocurrency exchanges, generating deepfake videos and forged documents after AI analysis of verification mechanisms. Google Threat Intelligence Group documented similar activity involving ARTEX, Hexstrike, and CAIRN, noting faster attack execution and automated privilege escalation against Active Directory environments.

安全客•AI Security

WDTA and CSA GCR Open Global Call for AI Security Experts to Build DigiBridge Talent Platform

The World Digital Technology Academy and Cloud Security Alliance Greater China Region have launched a public recruitment drive for certified instructors, course experts, case mentors, and project reviewers. The call supports the DigiBridge global digital talent development platform and its AI capability framework across five specialized tracks. Selected experts will deliver training, co-create industry cases, mentor learner projects, and participate in certification evaluations. The platform aims to close the digital and AI skills gap by connecting learners, universities, and enterprises through structured courses and practical projects. Applications close on November 18 at 18:00, with selected candidates joining a long-term global expert network.

Habr•AI Security

F6 Open-Sources MCP Server for XDR to Power AI Assistants in SOC Operations

F6 has released an open-source MCP server that connects AI assistants directly to its F6 XDR platform. The server enables tools such as OpenCode, Claude, Cursor, and Copilot to search alerts, incidents, emails, files, and host events, collect artifacts, and generate visualizations. Developed internally to reduce routine workload on SOC analysts, the solution also supports threat hunters, managers, and implementation engineers. The release includes twelve specific tools for querying, filtering, artifact handling, and alert management, with optional controls for sensitive operations. Real-world cases demonstrate its use in email statistics, alert triage, threat hunting from third-party reports, and analysis of a long-running miner-to-implant campaign. Cost and token usage data from DeepSeek V4.1 Flash runs are provided alongside practical recommendations for safe deployment.