Channel
@llmsecurity
On this record: Growth · Engagement · Reactions · Posts · Citations · Cite this entry
1,903subscribers
+11 since we began measuring on 6 August 2026
Risers and fallers across the register · movement among entries of 1,000–3,162.
| Telegram ID | -1002073963766 |
|---|---|
| Type | Channel |
| Username | @llmsecurity |
| Created | Between 1 November 2023 and 31 May 2024— estimated from Telegram’s id allocation, not measured. How this range is calculated. |
| First recorded | 6 August 2026 |
| Last confirmed live | 12 August 2026 |
| Measurements held | 4 |
| Confirmed unchanged | 1 time, most recently 12 August 2026 |
| On Telegram | t.me/llmsecurity |
| Measured (UTC) | Subscribers | Change |
|---|---|---|
| 12 Aug 2026, 22:42 | 1,903 | +4 |
| 9 Aug 2026, 20:31 | 1,899 | +6 |
| 7 Aug 2026, 04:36 | 1,893 | +1 |
| 6 Aug 2026, 12:03 | 1,892 | first reading |
14 posts held, back to 4 February 2026 — the reader has not yet reached the start of this channel’s public history, so older posts may sit further back, unread. Read across 2 pagesof Telegram’s post history, 20 posts per page.
ERR is average views per post over the last 30 days divided by subscribers, the definition TGStat uses, so this figure is comparable with the one you will see elsewhere. It falls structurally as a channel grows: a high ERR on a small channel and a low one on a large channel describe reach mathematics, not quality. We publish the figure and the sample it came from and pass no verdict on it.
ER is defined industry-wide as (forwards + reactions + comments) ÷ views— note the denominator is views, not subscribers. Telegram’s public web preview carries views and reactions but not forward or comment counts, so the reaction rate above is the reactions term only and is therefore a floor: the true ER for this channel is higher by an amount we have not measured and will not estimate.
| Window | Rolling 30 days · latest post in window 7 August 2026 |
|---|---|
| Posts held | 14 (4 February 2026 – 7 August 2026) |
| Views total | 2,160 |
| Reactions total | 27 |
| Forwards / comments | not exposed by the public surface — not measured, not estimated |
| Readings taken | 7 Aug 2026, 17:21 UTC |
Views are a single reading per post, taken at the time above. A post published in the last day or two is still accumulating views, which pulls the 30-day average down slightly. That is a property of the standard definition rather than a fault in it, so we keep the definition rather than “correcting” the number into something nobody can reproduce.
Precision. Telegram publishes view counts on its public widget in short form — 8.12K, 3.7M — so any reading at or above 1,000 reaches us rounded to three significant figures, and only counts below 1,000 are exact. Averages and rates derived from them are shown to the same precision rather than to the unit: a figure like 3,701,250 would assert digits nobody measured.
Reaction counts are published per emoji and rounded the same way, so a total below 1,000 is exact and a larger one is a sum that may carry a rounded component from each emoji above 1,000. Because it is a sum, it does not look rounded — read a large reaction total as three significant figures per contributing emoji rather than as the figure it prints.
106 reactions across 14 posts, in 6 distinct kinds. The most used accounts for 37.7% of them.
| Reaction | Count | Share | Share, drawn |
|---|---|---|---|
| 👍 | 40 | 37.7% | |
| 🥰 | 24 | 22.6% | |
| custom 5370711627026931838 | 20 | 18.9% | |
| 🦄 | 17 | 16.0% | |
| 🌚 | 4 | 3.77% | |
| 🥴 | 1 | 0.943% |
Custom emoji. One row above is a Telegram custom emoji, which the public preview renders as an element carrying only a numeric id — no character, and no image we can reach. The id is printed as-is rather than substituted with a look-alike glyph, because a stand-in would be our invention showing where a measurement should be. The count beside it isTelegram’s.
No sentiment is inferred, and none should be read in. This table is ordered by count and by nothing else. Emoji do not carry stable meaning across languages or communities — 🙏 is thanks in one channel and mourning in another — so we publish which ones were pressed and how often, and pass no judgement on what an audience meant by them.
Precision. Telegram publishes reaction counts per emoji and short-forms each one — 4.34K, 1.2M — so any single kind at or above 1,000 reaches us at three significant figures, and only counts below 1,000 are exact. The shares above are ratios of those figures and carry the same error. This is also why the total here can differ slightly from a reaction total printed elsewhere on the page: both are sums of the same rounded parts, taken over samples with different edges.
Coverage. Reactions were read on 14 of the 14 sampled posts in this sample. Summed by Telegram’s own count on each post — not by adding up the per-emoji breakdown above — those same posts carry 106reactions in total: the kind of figure the paragraph above means by “a reaction total printed elsewhere on the page”.
Measured over the 14 most recent posts we hold, published 4 February 2026 to 7 August 2026, using the newest reading held for each. Telegram Stars are excluded: they are a payment, not a reaction, and they have their own section.
Shieldstral Calvi et al., Mistral AI, 2026 Блог, статья, веса Французы из Mistral затюнили малыша Ministral-3B не 54 миллионах сэмплов и сделали из него еще одну гардрейл-модель – Shieldstral. В отличие от многих других подобных моделей, Shieldstral (как gpt-oss-safeguard) работает не с фиксированными категориями, а с задаваемыми пользователем политиками. При этом он является мультимодальным и, что важно, официально…
🦄2
Вы, возможно, помните историю с глитч-токенами – словами, которые соответствуют одному токену, который редко встречался в обучении и появление которого в промпте приводило к аномалиям – от невозможности повторить это слово до генерации моделью зловещих пророчеств. Они помогают понять, как смысл, хранимый LLM, зависит от геометрии векторного пространства, а в практическом смысле – помогают фингерпринтингу моделей или …
custom 53707116270269318389👍6
История об асимметрии в двух частях 16 июля компания Huggingface сообщает, что они столкнулись с кибератакой на свою инфраструктуру, которая была от начала до конца проведена силами LLM. В результате атаки были скопрометированы несколько датасетов и учетных записей. Никаких деструктивных действий, закладок в цепочке поставок – достаточно странно, неужели атакующих реально интересовали закрытые датасеты? Атака произо…
🥰8🌚2
Claude Code Is Steganographically Marking Requests Thereallo, 2026 Блог В сфере применения LLM есть большая проблема с доверием. Мы не можем на сто процентов доверять большим языковым моделям, так как они принимают решения вероятностно – отсюда все исследования на тему misalignment, scheming, sandbagging (намеренного занижения результатов оценок моделью), жульничества на бенчмарках и так далее. При этом мы предпола…
custom 53707116270269318386🌚1👍1
SearchLeak: How We Turned M365 Copilot Into a One-Click Data Exfiltration Weapon Dolev Taler, Varonis, 2026 Блог Еще один кейс непрямой промпт-инъекции с эксфильтрацией данных в Microsoft 365 Copilot, представленный компанией Varonis. Исследователи демонстрируют, как через компонент Enterprise Search можно украсть у клиента любые конфиденциальные данные, к которым есть доступ у ассистента. Как и во многих других по…
🥰3👍1
Представьте, что вы решили на своем скромном сайте для обмена квадратными фоточками с прикольными фильтрами сделать форму восстановления пароля, в которую нужно ввести логин и адрес почты, чтобы получить код восстановления. Вероятно, вы бы перед отправкой кода проверили, что данная почта действительно связана с этим аккаунтом. Может быть, в таком случае вы бы и не спрашивали логин. Возможно, вы бы использовали третий…
👍3🦄3
GLiNER Guard: Unified Encoder Family for Production LLM Safety and Privacy Bogdan Minko, Sabrina Sadiekh, Evgeniy Kokuykin, 2026 Препринт, блог, веса Большинство guardrail-моделей, которые выходят в опенсорс – это адаптации больших декодер-моделей, типа Qwen3 → Qwen3-Guard, в особо запущенных случаях — еще и ризонеров (привет, gpt-oss-safeguard-120B). Сегодня мы посмотрим на GliNER Guard — модель от HiveTraceLab, ко…
👍9
Главным событием на пересечении ИИ×ИБ в 2026 на текущий момент является переход LLM-агентами порога, за которым они становятся полезными для поиска уязвимостей. Даниэль Стенберг, мейнтейнер cURL, который в январе из-за вала слоп-репортов закрыл Bug Bounty, в апреле написал «The slop situation is not a problem anymore». Уровень тревоги в сообществе поддерживается маркетинговым департаментом Anthropic, размеренно выдаю…
🦄7👍4custom 53707116270269318383🥴1
A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models Kazemi et al., Apple, 2026 Препринт Все помнят abliteration (Arditi, 2024) — white-box метод снятия элайнмента, заключающийся в вычитании вектора из residual stream. Сегодня мы посмотрим на очень интересную статью, в которой исследователи из Apple демонстрируют, что расцензурирования модели можно достичь еще проще — изменив активации о…
🥰5👍4
Measuring AI Agents’ Progress on Multi-Step Cyber Attack Scenarios Folkerts et al., AISI, 2026 Блог, статья AISI, занимающаяся безопасностью ИИ в интересах правительства Великобритании, поделилась статьей об оценке способностей больших языковых моделей к кибератакам в сложных многошаговых сценариях – на киберполигонах. Более распространенным способом оценки наступательных способностей LLM являются CTF (как правило,…
🥰5custom 53707116270269318381
OWASP Agentic Skills Top 10 Сайт Если в прошлом году «кошмаром кибербезопасности» называли MCP-сервера, то теперь сна специалистов по ИИ-безопасности лишают навыки, или скиллы. По сути, скилл – это запакованная папка заданной структуры, содержащая основной промпт (SKILL.md), дополнительные подгружаемые инструкции, а также необходимые исполняемые файлы и ресурсы, например, данные. Кошмарность скиллам придают следующи…
👍7🥰3🌚1
Manipulating AI memory for profit: The rise of AI Recommendation Poisoning Microsoft Defender Security Research Team, 2026 Блог В позапрошлом году мы рассказывали на Offzone, как непрямая промпт-инъекция в документе может отравлять память ChatGPT, и предсказывали, что если раньше вы чистили компьютер родителей от браузерных тулбаров, сейчас – смартфон от оптимизаторов батарей, то в будущем будете очищать память LLM-…
👍1
Showing the 12 most recent of 14 posts we hold for @llmsecurity. View and reaction counts are the latest single reading for each post, not a live figure, and a recent post is still accumulating both. A view count marked ≈ was rounded by Telegram before we ever saw it — t.me prints views in full below 1,000 and to three significant figures above, so ≈1,200,000 means somewhere between 1,150,000 and 1,249,999. Unmarked counts are exact. Text is reproduced from the public post preview and truncated for length.
Citation-graph rank — 678,120 of 1,160,990entries in the measured graph. A weighted position computed from the forward and mention edges below — republished posts weigh more than named mentions — and recomputed periodically, over the whole graph. Published only as this ordinal position, never as a score: a position is a fact, and a score printed beside one channel’s name would read as a verdict this register does not make. The two counts beneath stay separate for the same reason mentions are never summed with forwards anywhere else on this page — a named-by count costs nothing to manufacture. The top 100 by this measure, or how it is computed.
Republished by
Channels on the register that have forwarded this channel's posts into their own feed.
Built only from forwarded posts we have actually read, on both sides. Coverage is early and deliberately incomplete: a missing link means we have not read the post that would prove it, never that the relationship does not exist. Counts are distinct forwarded posts observed, so they only ever go up as we read more.
Named by 2 registered channels — every channel on the register whose own posts have named this one, by its current username or any other username it currently holds, merged from two separately captured readings of the same fact so a namer caught by only one of them is not missed and a namer both caught is not counted twice. A username this channel has since dropped is not matched — that handle may belong to someone else now, and crediting today’s namer to yesterday’s owner would misattribute it.
Named by
Channels on the register whose posts name this channel's handle.
Names
Channels on the register whose handles appear in this channel's posts.
A mention is a weaker signal than a forward and is counted separately for that reason — naming a channel is not republishing it, and a handle in a post body is easy to place deliberately. The post counts beside each row below are distinct posts in which the handle appeared, from posts we have read on both sides — the “Named by N registered channels” figure above is a different count, of distinct NAMING CHANNELS rather than posts, and is not the sum of the rows under it.
A live page changes as we take new readings, so a citation should name the measurement it is based on, not just the URL. The line below cites the subscriber count as measured 12 August 2026 — this entry's latest reading, not the date you are reading this.
“llm security и каланы” (@llmsecurity), 1,903 subscribers as measured 12 August 2026. Telegram Register, tgregister.com/channel/llmsecurity.
Full measurement history, CC BY 4.0. Every reading this register holds for this entry, not just the latest one, as a dated, downloadable record: CSV · JSON. Free to use with attribution to tgregister.com. Each file carries its own generation timestamp, which is the figure to cite for exactly when the data was retrieved.