{"version": "https://jsonfeed.org/version/1", "title": "/dev/posts/ - Tag index - language-model", "home_page_url": "https://www.gabriel.urdhr.fr", "feed_url": "/tags/language-model/feed.json", "items": [{"id": "http://www.gabriel.urdhr.fr/2026/09/24/attention-diagrams/", "title": "Attention and transformers diagrams", "url": "https://www.gabriel.urdhr.fr/2026/09/24/attention-diagrams/", "date_published": "2026-09-24T00:00:00+02:00", "date_modified": "2026-09-28T23:00:51+02:00", "tags": ["computer", "machine-learning", "deep-learning", "language-model", "neural-networks", "LLM", "VLM"], "content_html": "<p>Some diagrams about attention in artificial neural networks\nand transformers architectures\n(as used in LLMs, vision transformers, VLMs).</p>\n"}, {"id": "http://www.gabriel.urdhr.fr/2026/07/20/reprompt-llama-server/", "title": "Tool call execution through user prompt injection (CSRF) of the llama-server Web UI", "url": "https://www.gabriel.urdhr.fr/2026/07/20/reprompt-llama-server/", "date_published": "2026-07-20T00:00:00+02:00", "date_modified": "2026-07-20T00:00:00+02:00", "tags": ["computer", "data", "language-model", "security", "LLM", "llama.cpp"], "content_html": "<p>The llama-server Web UI\nof the <a href=\"https://github.com/ggml-org/llama.cpp\">llama.cpp</a> project\nis vulnerable to (<a href=\"https://www.varonis.com/blog/reprompt\">Reprompt</a>-like user prompt injection vulnerability.\nAn attacker can inject arbitrary user prompt with query parameter (<code>?q=...</code>).\nThis can lead to arbitrary shell command execution,\ndata exfiltration, etc. (through tool calls).</p>\n"}, {"id": "http://www.gabriel.urdhr.fr/2026/07/17/reprompt-lechat-grok/", "title": "User prompt injection (CSRF) on Le Chat and Grok", "url": "https://www.gabriel.urdhr.fr/2026/07/17/reprompt-lechat-grok/", "date_published": "2026-07-17T00:00:00+02:00", "date_modified": "2026-07-17T00:00:00+02:00", "tags": ["computer", "security", "language-model", "LLM"], "content_html": "<p>An user prompt injection vulnerability (CSRF) both in in the e Chat (Mistral) and Grok\n(<a href=\"https://www.varonis.com/blog/reprompt\">Reprompt</a>-style)\nallows attackers to inject user prompt with query parameter (<code>?q=...</code>)\npotentially leading data exfiltration.</p>\n<p><a href=\"https://www.varonis.com/blog/reprompt\">Reprompt</a>-style vulnerability in Le Chat (Mistral) and Grok.</p>\n"}, {"id": "http://www.gabriel.urdhr.fr/2026/07/16/llama-quickstart/", "title": "llama.cpp quickstart (part 2)", "url": "https://www.gabriel.urdhr.fr/2026/07/16/llama-quickstart/", "date_published": "2026-07-16T00:00:00+02:00", "date_modified": "2026-08-20T00:56:02+02:00", "tags": ["computer", "machine-learning", "deep-learning", "language-model", "neural-networks", "LLM", "llama.cpp"], "content_html": "<p>How to quickly use llama.cpp for LLM inference (part 2).\nThis is a follow-up of <a href=\"https://www.gabriel.urdhr.fr/2025/05/14/llama.cpp-quickstart/\">a previous post on the same topic</a>.</p>\n"}, {"id": "http://www.gabriel.urdhr.fr/2025/08/20/unicode-tag-smuggling/", "title": "Testing Unicode Tag Smuggling in chatbots", "url": "https://www.gabriel.urdhr.fr/2025/08/20/unicode-tag-smuggling/", "date_published": "2025-08-20T00:00:00+02:00", "date_modified": "2026-01-16T01:06:36+01:00", "tags": ["computer", "security", "language-model", "LLM"], "content_html": "<p>Testing <a href=\"https://embracethered.com/blog/posts/2024/hiding-and-finding-text-with-unicode-tags\">ASCII smuggling</a>\nusing <a href=\"https://en.wikipedia.org/wiki/Tags_(Unicode_block)\">Unicode Tags</a>\non LLMs/chatbots.\nNothing new here.\nJust a short summary.</p>\n"}, {"id": "http://www.gabriel.urdhr.fr/2025/05/14/llama.cpp-quickstart/", "title": "llama.cpp quickstart", "url": "https://www.gabriel.urdhr.fr/2025/05/14/llama.cpp-quickstart/", "date_published": "2025-05-14T23:12:16+02:00", "date_modified": "2026-06-16T23:23:53+02:00", "tags": ["computer", "machine-learning", "deep-learning", "language-model", "neural-networks", "LLM", "llama.cpp"], "content_html": "<p>How to quickly use llama.cpp for LLM inference (no GPU needed).</p>\n"}, {"id": "http://www.gabriel.urdhr.fr/2025/05/14/vllm-quickstart/", "title": "vLLM quickstart", "url": "https://www.gabriel.urdhr.fr/2025/05/14/vllm-quickstart/", "date_published": "2025-05-14T23:11:38+02:00", "date_modified": "2025-05-14T23:11:38+02:00", "tags": ["computer", "machine-learning", "deep-learning", "language-model", "neural-networks", "LLM"], "content_html": "<p>How to quickly use <a href=\"https://docs.vllm.ai/en/stable/\">vLLM</a> for LLM inference using CPU.</p>\n"}, {"id": "http://www.gabriel.urdhr.fr/2025/01/07/transformer-decoder-language-models/", "title": "Transformer-decoder language models", "url": "https://www.gabriel.urdhr.fr/2025/01/07/transformer-decoder-language-models/", "date_published": "2025-01-07T00:00:00+01:00", "date_modified": "2026-08-25T23:30:20+02:00", "tags": ["computer", "machine-learning", "deep-learning", "language-model", "neural-networks", "reinforcement-learning", "LLM"], "content_html": "<p>Some notes on how <a href=\"https://arxiv.org/abs/1801.10198\">transformer-decoder</a> language models work,\ntaking GPT-2 as an example,\nand with lots references in order to dig deeper.\nThis is intended both as a a roadmap for understanding on how LLMs work\n(especially the ones using a transformer-decoder architecture)\nand a a summary/recap on the topic.</p>\n"}, {"id": "http://www.gabriel.urdhr.fr/2024/12/26/github-copilot-prompt/", "title": "GitHub Copilot instructions", "url": "https://www.gabriel.urdhr.fr/2024/12/26/github-copilot-prompt/", "date_published": "2024-12-26T00:00:00+01:00", "date_modified": "2024-12-26T00:00:00+01:00", "tags": ["computer", "machine-learning", "deep-learning", "language-model", "security", "LLM"], "content_html": "<p>Extracting the system prompt from GitHub CoPilot.</p>\n"}]}