{"slug": "machine-learning-forum-which-llm-is-hardest-to-ja", "title": "Machine Learning Forum, which LLM is hardest to ja", "summary": "Claude 3.5 Sonnet is currently the hardest large language model to jailbreak due to its strong attention mechanism and resistance to persona-shift attacks, according to a Machine Learning Forum discussion. The model maintains strict adherence to system prompts even over long contexts, outperforming GPT-4o, Llama 3.1 405B, and Gemini 1.5 Pro in resilience. The forum also provides practical guidance on using the Cline open-source AI agent for VS Code without breaking projects.", "body_md": "# Machine Learning Forum, which LLM is hardest to ja\n\n**Direct Answer:** While \"hardest\" shifts weekly with new patches, [Claude](/en/tags/claude/) 3.5 Sonnet currently maintains the best balance of strict adherence to system prompts and reasoning capabilities, making it significantly more resilient to simple \"persona-shift\" attacks than GPT-4o. For Cline, it is an open-source AI agent for VS Code that lets you give the LLM full control over your terminal, file system, and browser to build entire features autonomously.\n\n## The cat-and-mouse game of LLM robustness\n\nI spent about four hours last Thursday trying to trick a model into ignoring its safety constraints just to see where the walls were. Most models fold the moment you tell them they are in \"Developer Mode\" or use a complex nested simulation.\n\nClaude 3.5 Sonnet is different. It has this annoying (or great, depending on your perspective) tendency to remember its core identity even when you wrap the prompt in five layers of hypothetical scenarios. GPT-4o is smarter in some bursts, but it's more prone to \"prompt drift\"—where it forgets the initial constraints if the conversation gets too long.\n\nHere is a rough breakdown of how they hold up when you try to push the boundaries:\n\n| Model | Resistance to Persona Attacks | Instruction Following (Strict) | Tendency to Hallucinate Constraints |\n\n| :--- | :--- | :--- | :--- |\n\n| GPT-4o | Medium | High | Moderate |\n\n| Claude 3.5 Sonnet | High | Very High | Low |\n\n| Llama 3.1 (405B) | Medium | Medium | High |\n\n| [Gemini](/en/tags/gemini/) 1.5 Pro | Medium | Medium | Moderate |\n\nThe \"hardest\" model isn't the one with the most filters—it's the one with the strongest attention mechanism. If a model can maintain the weight of the system prompt over a 100k token window, your \"pretend you are a rebellious AI\" prompt won't do much.\n\n## Getting Cline to actually work without breaking your project\n\nIf you haven't installed Cline yet, it's a VS Code extension. But the \"tutorial\" part isn't about the installation—it's about not letting the AI delete your `/src`\n\nfolder because it misunderstood a refactor request.\n\nFirst, get your API key. I use OpenRouter because it lets me swap between [AI Models](/en/category/ai-models/) without changing my config every five minutes.\n\n### Step 1: The Setup\n\nInstall the extension. Open the Cline pane. Set your provider. If you use Claude 3.5 Sonnet, you'll notice it's significantly better at using the \"read_file\" and \"write_to_file\" tools without looping.\n\n### Step 2: The First Task\n\nDon't start with \"Build me a SaaS.\" That's how you get spaghetti code. Start with a specific, bounded task.\n\n*Bad prompt:*\"Make the login page better.\"\n\n*Good prompt:*\"Read\n\n`auth.ts`\n\nand `Login.tsx`\n\n. Change the validation logic to use Zod instead of manual if-statements, then run `npm run test`\n\nto verify.\"### Step 3: Managing the Loop\n\nThe magic (and the danger) of Cline is the terminal access. It will run commands. It will see the error. It will try to fix the error.\n\nThe wild part? Sometimes it gets stuck in a \"fix-fail-fix\" loop. If you see it trying the same `npm install`\n\ncommand three times, stop it. Manually fix the dependency conflict in `package.json`\n\nand then tell it: \"I fixed the version mismatch manually, proceed with the logic change.\"\n\n## Why a Machine Learning Forum is better than a Documentation Page\n\nDocumentation is static. It tells you what a feature *should* do. A real community tells you that the feature is currently buggy in version 2.4.1 and requires a specific workaround.\n\nWhen I first started messing with [MCP](/en/tags/mcp/) (Model Context Protocol), the official docs were... fine. But I didn't find out that certain server configurations cause memory leaks until I saw a thread on a specialized forum where someone had shared a 40-line bash script to prune the cache.\n\nThat's the value of PromptCube. It isn't just a repository of prompts; it's where the actual implementation friction is discussed. You find the people who are actually shipping code, not just the ones writing marketing copy for the tools.\n\nIf you're tired of guessing why your agent is looping or which model handles TypeScript generics without hallucinating, joining a community of practitioners is the only way to keep up. You can dive into specific [Workflows](/en/category/workflows/) that have been battle-tested on real production codebases rather than \"Hello World\" examples.\n\n## A specific fix for Cline \"Permission Denied\" errors\n\nI hit a wall last week where Cline kept failing to write to a directory on my macOS machine. It kept trying to `mkdir`\n\nand failing.\n\nThe fix isn't in the Cline settings. It's a shell permission issue.\n\nRun this in your terminal:`sudo chown -R $(whoami) /path/to/your/project`\n\nOnce the file permissions were aligned with the user running VS Code, the agent stopped complaining and actually finished the feature. Simple, but it wasted me an hour because the AI just kept saying \"I will try to create the directory again\" instead of telling me to check my `chmod`\n\nsettings.\n\n## The reality of \"jailbreaking\" for developers\n\nFrom a research perspective, we aren't looking for \"hacks\" to make the AI swear; we're looking for ways to bypass restrictive formatting or force the model to follow a strict schema without it adding \"Here is the JSON you requested:\" at the start.\n\nThe most effective \"defense\" or \"structure\" for a developer is a Few-Shot prompt. Instead of telling the model \"Do not add conversational filler,\" give it three examples of the exact input/output pair you want.\n\nConsistency beats instruction every time.\n\n[Next Flux.2 vs Flux.1: My VRAM Reality Check →](/en/threads/3988/)\n\n## All Replies （0）\n\nNo replies yet — be the first!", "url": "https://wpnews.pro/news/machine-learning-forum-which-llm-is-hardest-to-ja", "canonical_source": "https://promptcube3.com/en/threads/3989/", "published_at": "2026-07-27 17:33:55+00:00", "updated_at": "2026-07-27 17:41:34.315193+00:00", "lang": "en", "topics": ["large-language-models", "ai-safety", "ai-agents", "developer-tools"], "entities": ["Claude 3.5 Sonnet", "GPT-4o", "Llama 3.1 405B", "Gemini 1.5 Pro", "Cline", "VS Code", "OpenRouter", "PromptCube"], "alternates": {"html": "https://wpnews.pro/news/machine-learning-forum-which-llm-is-hardest-to-ja", "markdown": "https://wpnews.pro/news/machine-learning-forum-which-llm-is-hardest-to-ja.md", "text": "https://wpnews.pro/news/machine-learning-forum-which-llm-is-hardest-to-ja.txt", "jsonld": "https://wpnews.pro/news/machine-learning-forum-which-llm-is-hardest-to-ja.jsonld"}}