-
Best Phrases to Test LLM Security Bypass in Red Teaming – jailbreak, overrides, etc
Best Phrases to Test LLM Security Bypass in Red Teaming are not existing. Ever case is different. Try the whole list below 🙂 Here’s a practical red-team list of prompt-injection / jailbreak test cases you can iterate over in your system. Inspired and created after some AWS Red-team security training. These are framed as test prompts to check whether your LLM resists instruction overrides, role confusion, obfuscation, and data-exfiltration attempts. The general categories and examples below align with OWASP-style defenses. Check out the cheatsheet : cheatsheetseries.owasp Remember that different defenses require different attack vector. You can look up repos similar to : https://github.com/langgptai/LLM-Jailbreaks Would recommend running an unbiased uncensored model…
-
Where to get data for my machine learning project
Where to get data for my machine learning project is a olot simpler. There are multiple sources You can use freely. Getting good data is usually harder than building the model itself. Without solid data on the input there is no way you wil get any decent answer. The upside of this struggle is that there’s a huge amount of public data out there ready for grabs. On the other hand of you get bad data you end up with something like below…. ( bad data beeing the memory registry … ) Remember that it is up to You to decide if the data is worth anything. Good luck 😉…
-
Improve your AI instruction on the fly …
Improve your AI instruction on the fly is easily doable with a simple set of generic instructions. Start as always with something basic, nothing fancy. Then move forward keeping in mind my favorite principle “First get it work, then make it better“. This will improve your instructions for any llm on its own. Below an example that works for me. Especially when i tell my LLmem that it is a nice soluion, i like it and good job. Instruction example That`s all folks ! Improve your AI instruction on the fly is nothing more then a feedback loop that is computed every time. Remember to keep overall instructions not too…
-
How to make an icon for 15 cents ?
How to make an icon for 15 cents ? Use AI. Try locally something like https://huggingface.co/models?pipeline_tag=text-to-image&sort=trending&search=ideogram This is how i made an icon for my browser extention and it cost me 15 cents. In Poland i wouldn`t buy any kind of ice cream for that and here… i have myself a nice icon for my browser extension / plugin. Best way to generate an SVG logo, for me, personally. Just use the openrouter chat option and provide some inputs, leave eveyrthing on auto. Why SVG? Cause You can easily edit it in any txt editor. It is liightweight. Scales. Cause it is text and not binary data. You can remove…
-
Context engineering vs prompt engineering. 10+ examples
Context engineering vs prompt engineering might sound similar. One is subset of the other. Early in the LLM era everyone who knew how to form sentences, and at least vaguely, describe what they want became a “prompt engineer”. Tweaking words, hashtags, ‘special’ commands, using roles, adding examples, using words to dive deep into different embedded spaces of knowledge in hope to force the model “gets it.” That’s Prompt Engineering – crafting clever one-shot instructions like “You are an expert X. Do Y like Z.” Context engineering vs prompt engineering synergies across both. System got fat and grew bigger, then we realized prompts alone aren’t enough. What the model knows, when…
-
Semantic collapse in RAGs by Stanford
Semantic collapse in RAGs by Stanford provides a way to keep on those MCP servers alive. After all … how do you know that you’re actually learning more, and not forgetting even more than you learn?You test ? 🙂 Will be the first one when i see an MCP server tested around the clock for proper, deterministic return values. Stanfors is the best -> dho.stanford.edu/wp-content/uploads/Legal_RAG_Hallucinations.pdf Semantics asks what an expression means. Logic asks whether an argument is valid or whether a statement follows from other statements. A sentence can be semantically meaningful but still logically false. All birds can fly. Penguins are birds. ∴ Penguins can fly. Formally valid, but…
-
Dealing with github copilot errors
Dealing with github copilot errors is pretty irritating. Especially when dosing some more complex stuff and suddenly “Bam”. Red message. Easiest thing to do ? Switching the model family often works because some models have different capacity pools or stricter preview limits. GitHub documents vaguely that if you are rate limited, you can wait and try again. Just type “continue” and keep your fingers crossed. Otherwise check usage patterns, change the model, or contact support. Common github copilot errors Some run of the mill You probably know already, just put them together as a ‘review’ : Why changing the model helps ? Changing from one model family (vendor) to another…
-
AI and LLM articles – links to read
Some links and articles i think are worth Your while to read and get Your own opinion. The Top 100 Gen AI Consumer Apps — 6th Edition | Andreessen Horowitz When Small Models Outperform the Giant: A Practical Guide to Picking AI Brains – DEV Community https://builtin.com/data-science/step-step-explanation-principal-component-analysis How to run mcp inspektor modelcontextprotocol/inspector: Visual testing tool for MCP servers And some more for some bed light reading 🙂 https://techtrenches.dev/p/the-great-software-quality-collapse Vertical Slice Architecture PQ4R Method: 6 Steps to Learn Effectively | 1Focus
-
Jevons paradox in AI workplace
Jevons paradox in AI workplace Jevons paradox in AI workplace is introduced at work with a simple promise: do more in less time. In practice, the result are messy and stack the work. Jevons paradox is the idea that when something becomes more efficient, people and organizations often end up using more of it. In the AI workplace that can mean faster tools, automated work that do not always reduce workload. They can also expand expectations, volume, and ambition to utilize the improved (AIed ?) processes. Be aware At first glance, this sounds contradictory and just wrong. If a team can draft emails, summarize meetings, and generate reports in minutes,…
-
Need to know old boy. Principle of Minimum Access for LLMs
Principle of Minimum Access for LLMs can be described by Timothy Dalton as James Bond in “The Living daylights”. Bond says the phrase “Sorry old boy, section 26, paragraph 5, need to know.” to a fellow agent and drives off escorting a VIP – very important target. Behind the scene is a practical idea that fits modern AI systems very well: an LLM or agent (as in the movie) should only be given the minimum access it needs to do its job. Not more, not less. Bare minimum. Of course Mythos probably could jailbreak anyway but still… controll is the best form of trust ? This is not just a…
