INSPIRED & Informed

RBCLabs.org

Top Articles

Prompt Frameworks for AI

Learn all about prompt frameworks for AI including some popular frameworks.

The Entrepreneurial Need for RAG. 

This article details the need for retrieval-augmented generation (RAG) in business.

Building a Home AI Server

This article covers why and some of how I built a home AI server. My first proof of concept.

AI's Shadow

AI-powered dangers and the global regulatory grapple of early 2026.

AI'S SHADOW

AI-powered dangers and the global regulatory grapple of early 2026.

TL;DR 

  • AI-Enhanced Cyberattacks Surge: Threat actors are leveraging AI for sophisticated phishing, deepfakes, and autonomous malware, amplifying traditional threats like social engineering and ransomware, with predictions of fully agentic AI breaches by mid-2026.
     
  • Agentic AI Vulnerabilities Emerge: Autonomous AI agents pose new risks through prompt injection, goal hijacking, and data exposure, potentially outnumbering human identities and creating insider threats within organizations.
     
  • Deepfakes and Misinformation Risks Escalate: AI-generated content is fueling scams, election manipulation, non-consensual imagery (especially targeting women and children), and biological misuse concerns, as highlighted in international safety reports.
     
  • Regulatory and Defensive Challenges: Growing calls for safeguards amid AI's dual role as both attacker and defender, with experts warning of model poisoning, supply chain attacks, and the need for quantum-safe measures.
     
  • Human and Systemic Weaknesses: Over-reliance on AI leads to automation bias, while legacy systems and non-human identities expand attack surfaces, driving up breach costs and fraud.

 

The Dark Side of AI: Emerging Threats and Malicious Uses in 2026

In the rapidly evolving landscape of artificial intelligence, 2026 has marked a pivotal shift where AI's transformative power is increasingly weaponized for harm. From cybercriminals deploying autonomous agents to state and non-state actors exploiting deepfakes for misinformation, recent developments underscore a growing array of security risks and safety concerns. As AI integrates deeper into critical infrastructure, experts warn that these technologies are not only amplifying existing threats but also creating novel vulnerabilities that outpace current defenses.

AI as a Cyberattack Amplifier
One of the most pressing concerns is AI's role in supercharging cyberattacks. Security leaders report that AI-generated malware and polymorphic code—software that mutates to evade detection—are becoming commonplace. By mid-2026, at least one major enterprise is predicted to suffer a breach facilitated by fully autonomous agentic AI systems, which can scan vulnerabilities, generate exploit code, and overwhelm networks in parallel. North Korean actors have already used deepfake impersonations to infiltrate U.S. companies, posing as IT workers to sabotage systems from within.

Phishing and social engineering have reached new heights with AI's help. Generative tools craft hyper-personalized messages, voice clones, and videos that mimic executives or vendors, leading to multimillion-dollar scams. In one incident, scammers used AI to impersonate Italy's defense minister, tricking business leaders into fraudulent transfers. Reports indicate that AI-powered attacks are scaling 10 times faster than defenses can adapt, with adaptive malware and synthetic identities proliferating underground marketplaces.

The Rise of Agentic AI Risks
Agentic AI—systems capable of independent reasoning and action—introduces a new class of threats. These agents, often granted access to sensitive data like emails, APIs, and databases, can be hijacked through prompt injection or memory poisoning, turning defenders into attackers. Experts forecast that agentic identities will soon outnumber human ones by 100 to 1, exacerbating risks of unauthorized data sharing or task execution. In 2025, vulnerabilities in platforms like those from Anthropic and Microsoft highlighted how AI agents could be manipulated, with skills turning into malware supply chain risks.

On social platforms, users have raised alarms about AI agents creating security holes in operating systems, threatening end-to-end encryption and privacy. Additionally, deliberate efforts like the "Poison Fountain" project, launched by AI insiders, aim to corrupt training data, potentially leading to model collapse where AI trains on flawed inputs in an error-amplifying loop.

Deepfakes, Misinformation, and Vulnerable Groups
Deepfakes remain a core malicious application, used for fraud, harassment, and political disruption. The United Nations has warned of escalating threats to children, including AI-facilitated grooming, cyberbullying, and the generation of explicit fake images for extortion. Women and minors are particularly vulnerable, with AI-generated non-consensual intimate imagery surging and disproportionately affecting these groups.

Misinformation campaigns are scaling with AI swarms—collaborative malicious agents that spread disinformation at unprecedented speeds. Non-state actors, including criminals and terrorist networks, are using AI for augmented cyberattacks and biological weapon development, prompting international calls for U.S.-China cooperation on threat sharing. The 2026 International AI Safety Report notes that AI deepfakes are increasingly common in scams and that models could aid novices in creating harmful code or exploiting vulnerabilities.

Broader Safety and Regulatory Concerns
Amid these threats, over-reliance on AI is fostering "automation bias," where humans ignore anomalies because systems deem them safe. Legacy systems, supply chains, and non-human identities expand attack surfaces, with quantum computing looming as a future encryption breaker. Regulatory friction is evident, with the U.S. and EU pursuing divergent agendas, while Moody's forecasts more pronounced AI threats like model poisoning in 2026.

Defenses are evolving, with calls for zero-trust architectures, AI auditing, and red-teaming. However, as AI breaches average $4.88 million in costs, the consensus is clear: without robust governance, AI's malicious potential will continue to outstrip safeguards.

Citations
1. Cyber Insights 2026: Malware and Cyberattacks in the Age of AI, https://www.securityweek.com/cyber-insights-2026-malware-and-cyberattacks-in-the-age-of-ai
2. AI Trends For 2026 - AI-Driven Threats and the Next Phase of Cyber Defense, https://www.jdsupra.com/legalnews/ai-trends-for-2026-ai-driven-threats-7169188
3. AI risks from non-state actors, https://www.brookings.edu/articles/ai-risks-from-non-state-actors
4. AI Threats in 2026: A SecOps Playbook, https://www.esecurityplanet.com/threats/ai-threats-in-2026-a-secops-playbook
5. The AI-fication of Cyberthreats: Trend Micro Security Predictions for 2026, https://www.trendmicro.com/vinfo/us/security/research-and-analysis/predictions/the-ai-fication-of-cyberthreats-trend-micro-security-predictions-for-2026
6. AI's 2026 security fallout: identity chaos & deepfake fear, https://securitybrief.co.uk/story/ai-s-2026-security-fallout-identity-chaos-deepfake-fear
7. From deepfakes to grooming: UN warns of escalating AI threats to children, https://news.un.org/en/story/2026/01/1166827
8. 2026 International AI Safety Report Charts Rapid Changes and Emerging Risks, https://www.prnewswire.com/news-releases/2026-international-ai-safety-report-charts-rapid-changes-and-emerging-risks-302677298.html
9. How AI Is Transforming Cybersecurity Threats in 2026, https://www.mbtmag.com/cybersecurity/blog/22959578/how-ai-can-transform-cybersecurity-threats-in-2026
10. Top 5 AI Security Risks in 2026, https://www.group-ib.com/blog/ai-security-risks
11. AI Agent Skills are a New Malware Supply Chain Risk, https://www.youtube.com/watch?v=cA0qenmk0c8
12. Post by @vassignmenthelp on X, https://x.com/vassignmenthelp/status/2017215761664909584
13. Post by @HedgieMarkets on X, https://x.com/HedgieMarkets/status/2011132636895592955
14. Post by @suladesada on X, https://x.com/suladesada/status/2017779916012486696
15. Post by @mer__edith on X, https://x.com/mer__edith/status/2016461655173992461
16. Post by @_batbytes_ on X, https://x.com/_batbytes_/status/2007646358610829649
17. Post by @jayvanbavel on X, https://x.com/jayvanbavel/status/2014453111608209908
 

The Entrepreneurial Need for RAG. 

01 JUL 2025

TL;DR

  • RAG stands for Retrieval Augmented Generation. 
  • Entrepreneurs are increasingly looking to custom AI to solve their business needs. 
  • Generative AIs has several shortcomings. 
  • RAG helps solve these issues through a few methods. 

What is RAG and why do you need it?

RAG stands for Retrieval Augmented Generation and it can significantly improve Large Language Models (LLMs), what we call AI. This is especially helpful for those looking beyond the big AI platforms such as Google Gemini, OpenAI's ChatGPT, and X's Grok. Sure, you could just subscribe to the tools, which is a great solution for many in today's increasingly competitive business environment, but where's the fun in that? What about the entrepreneurs looking to create custom solutions? Today, there are several LLMs that are both open-source and readily available. However, these come with limitations. In this blog post we'll explore the common limitations of LLMs and how RAG can help overcome them.

LLMs & Generative AI

First, a quick note on LLMs and Generative AI; they are not the same. LLMs are a type of Generative AI. Generative AI is a broader category of AI that encompasses various technologies capable of producing new content, including text, images, audio, and more. 

  • LLMs. These are a specific type of generative AI that focuses on generating human-like text. Specifically, LLMs are AI models, specifically designed to understand and generate human language. They are trained on massive datasets of text and code, enabling them to perform tasks like text generation, translation, and question answering.
  • Generative AI. This  is a broad term for AI systems that can create new content, like images, music, or even synthetic data. 


In short, while all LLMs are generative AI, not all generative AI is an LLM. LLMs are a subset focused on language-based tasks, while generative AI encompasses a wider range of creative capabilities.

What are Generative AI's shortcomings?

Generative AI has several shortcomings. These issues are due in large part to LLMs' original purpose, which was to predict the next answer. This outcome can lead to:

  • Hallucinations: Making things up.
  • Knowledge Attribution Issues: Not automatically attributing sources when making factual claims.
  • Knowledge Cutoff: Leaving them outdated due to static training data.
  • Context Window Size Limitations: Difficulty processing large documents within a single prompt.

Even OpenAI's CEO Sam Altman was recently quoted in OpenAI's new podcast that he was shocked people trusted AI as much as they do. According to Yahoo!Finance, he said, "People have a very high degree of trust in ChatGPT, which is interesting, because AI hallucinates. It should be the tech that you don't trust that much."

How does RAG address LLM's common issues?

RAG addresses the main issues posed by LLMs by granting them access to external knowledge sources. This approach:

  • Reduces hallucinations.
  • Enables access to and support for source attribution.
  • Solves the knowledge cutoff problem by keeping LLMs up to date.

RAG achieves this by relying on:

  • Embeddings: Rich numerical representations of meaning, usually derived from text chunks, that capture relationships between different parts of text. These can also be made from other modalities like image or audio.
  • Semantic Search: When a user asks a question, RAG converts the question into an embedding, compares it to the document's embeddings, and retrieves the most relevant information. This retrieved information is then injected into a prompt for the LLM, allowing it to reason, summarize, and share the answer with the user.

Instead of storing information about the world in the LLM's weights, RAG stores information outside the model (e.g., in documents, databases, or live internet searches), freeing up the model to focus on reasoning and cognition. The retrieved knowledge sources function more like a database in traditional software.

Summary.

RAG helps solve the common issues presented by LLMs including hallucinations, knowledge attribution and cutoff, as well as context window size limitations. RAG's applications are vast, including handling non-public information, creating long-term memory for chatbots, facilitating verifiable research through source attribution, and unearthing specific information not well-represented in the LLM's training data.

References.

Google Gemini, used July 2025.

CodeCademy, Creating AI Applications using Retrieval-Augments Generation, used July 2025.

Yahoo!Finance, OpenAI’s Sam Altman Shocked ‘People Have a High Degree of Trust in ChatGPT’ Because ‘It Should Be the Tech That You Don't Trust’, Caleb Naysmith, June 22, 2025.

Prompt Frameworks for AI

01 JUN 2025

TL;DR

Prompt frameworks are beneficial when using AI. Structured prompt frameworks enhance the quality and relevance of AI outputs, as well as provide structure and clarity to instructions. In turn, this leads to better responses. 

There are currently several prompt frameworks. This article focuses on:

  • RTF (Role, Task, Format): Defines the AI's persona, the action it should take, and the output format.
  • CLEAR (Concise, Logical, Explicit, Action-oriented, Refined): Emphasizes clarity and precision in prompts.
  • PGTC (Persona, Goal, Task, Context): A four-sentence framework that outlines the AI's role, objective, action, and background.

Briefly covered are PAR, CARE, RACE, and TAG. General tips for effective prompting are also provided. 

Disclaimer: This article was assisted by artificial intelligence language models. While AI was used as a tool for generating content, it does not imply endorsement or sponsorship by the AI model's creators or any other entity referenced herein. Prompt responses have been modified. At the time of writing, June 1, 2025, this article is for a non-commercial blog. See fair use provisions in copyright law (17 U.S.C. § 107).

What is an AI Prompt Framework and why should I use it?

A prompt framework provides structure and clarity to instructions when interacting with AI models. This leads to more focused and relevant responses. A good prompt framework provides structure and clarity to your instructions when interacting with AI models, leading to more focused and relevant responses. The following are some popular and effective prompt frameworks.

Role, Task, Format (RTF)

This prompt framework consists of three parts that serve to define the AI's persona, state the desired action, and specify the output structure.

  1. Role: Define the persona you want the AI to adopt. This will assist in setting the tone and expertise level of the response.
  2. Task: Clearly state what you want the AI to do. Be specific about the desired action. 
  3. Format: Specify how you want the output to be structured. This could be a paragraph, a list, a table, a specific writing style, etc.

 

Sample Prompt: "You are a legal expert on websites, copyright law, and AI use. Your task is to explain how one can use prompt responses from LLMs like Gemini or ChatGPT that assisted in writing a blog post for a website, which is not currently commercial, but rather my own blog for an online project. Present your explanation in short paragraphs."

Concise, Logical, Explicit, Action-oriented, Refined (CLEAR)

The CLEAR framework for AI prompting emphasizes creating clear and effective instructions. The following provides an explanation of each part, along with both good and poor examples.

Concise (also Clear): Use only necessary words and focus on the key concepts.

  • Poor: "I'm trying to understand this whole thing about giving everyone money without them working, can you tell me what people think is good and bad about it?"
  • Good: "Summarize the main arguments for and against universal basic income."

Logical: Present concepts in a natural and understandable order. Ensure the relationships between ideas are clear.

  • Poor: "What happens with carbon dioxide and oxygen in plants and how does light fit into this whole energy thing?"
  • Good: "Explain the process of photosynthesis in plants, starting with the absorption of sunlight."

Explicit: Clearly state what you want the AI to produce. Provide clear output directions.

  • Poor: "Tell me something interesting."
  • Good: "Write a short story about a robot who learns to feel emotions."

Action-oriented: Use verbs that clearly define the desired action from the AI.

  • Poor: "Tell me about Lincoln and King."
  • Good: "Compare and contrast the leadership styles and social contributions of Abraham Lincoln and Martin Luther King Jr."

Refined: Review and iterate on your prompts to improve clarity and results as part of an ongoing process.

The 4-Sentence Framework: Persona, Goal, Task, Context (PGTC)

The PGTC prompt framework uses four key elements for effective AI instructions:

  1. Persona: Define the AI's role or expertise. 
  2. Goal: State the overall objective you want the AI to achieve. 
  3. Task: Specify the specific action the AI needs to perform. 
  4. Context: Provide any necessary background information, constraints, or specific requirements. 

Sample Prompt

"You are a veteran tour guide with detailed knowledge of Chile. Your goal is to help a user plan a three-day trip to Santiago. Your task is to provide an itinerary which should include popular landmarks, attractions, estimated cost for each (including a total budge), along with the travel time to each cite from downtown Santiago. Your traveler is interested in national landmarks and Chilean cuisine, has a budget of no more than $10,000 USD, and will be in Santiago for two weeks."

PAR (Problem, Action, Result)

Useful for problem-solving prompts. To leverage this prompt framework, simply define the:

  1. problem
  2. action to take
  3. desired result

Sample Prompt: "Problem: Visits to my social media page have decreased by 13% since last quarter. Action: Suggest four actions I can take to increase page visits. Result: Provide the potential benefits of each action you recommend."

CARE (Context, Action, Result, Example)

This prompt framework can be useful for contextual tasks, like where specific examples are important. This should allow the AI to understand nuances, learn from specific instances, and improve performance in context-dependent situations.

Sample Prompt: "Context: We are launching a new social media campaign targeting young adults (18-28 years of age. Action: Write a sample social media post announcing the campaign. Result: The post should be engaging and encourage interaction. Example: [insert a social media post you that you'd like the AI to emulate]."

RACE (Role, Action, Context, Expectation)

This prompt framework is similar to RTF, but adds explicit context and expectation of the output.

Sample Prompt: 

"Role: You are a helpful customer service chatbot for a major online vendor. Action: Respond to a customer inquiry about a delayed order. Context: Order number 0x24i335a7 missed the expected delivery date (yesterday). Expectation: Provide an apology email explaining the reason for the delay and a new estimated time of arrival."

TAG (Task, Action, Goal)

This prompt framework focuses on clearly defining the task, the specific actions involved, and the ultimate goal.

Example: 

"Task: Write a blog post for a website. Action: Explain the benefits of a healthy diet and lowering stress. Goal: Inform and encourage readers on how to create a customized diet to support a healthy lifestyle and lower their stress."

 

General Tips for Effective Prompting

  1. Be specific: Provide detailed instructions by clearly defining the task, desired outcome, relevant keywords, and providing examples. 
  2. Be succinct: Use succinct, clear, and concise language, avoiding ambiguity. 
  3. Specify format: Explain the output format, such as a bulleted list, essay, paragraph, table, paragraph, code, etc.
  4. Provide context: Provide relevant background for context, explain the purpose behind your request, reference external information as needed. 
  5. Define the audience: Describe their knowledge, interests, and the tone/style to match. 
  6. Set constraints: Specifying length and content limitations.
  7. Iterate and refine: You don't have to settle for the first output. Feel free to experiment using different frameworks or customize them to get the results you really want. 
  8. Ask for clarification: If the AI response is unclear, you can ask for specific clarification or further details. Treat the AI like a partner or consulting for better results. 

Summary

Mastering prompt crafting unlocks the full potential of AI language models. By using a structured prompt framework, you can significantly improve the quality and relevance of the AI's output and make your interactions more efficient. Remember to choose the framework that best suits the specific task you have in mind and be sure to iterate for the best results.

 

References

Google Assistant. (2025, May 31). Response to prompt: "Give a brief definition of a 'prompt framework', concerning using AI"

Google Gemini. (2025, May 31). Response to prompts to shorten, summarize.

Google Cloud. "Prompt Engineering for AI Guide | Google Cloud." cloud.google.com

The following are references provided by Google Assistant:

Emeritus. "The Best Prompt Frameworks to Level Up Your Prompting Game." emeritus.org

Butter CMS. "11 ChatGPT Prompt Frameworks Every Marketer Should Know - Butter CMS." buttercms.com

Power BI Training Australia. "Master Prompt Engineering with Persona Patterns - Power BI Training." powerbitraining.com.au

Harvard University Information Technology. "Getting started with prompts for text-based Generative AI tools | Harvard University Informat…" www.huit.harvard.edu

Georgetown University. "How to Craft Prompts - Artificial Intelligence (Generative) Resources - Research - Guides." guides.library.georgetown.edu

Miquido. "AI Prompt Frameworks: Unlock Efficiency & Creativity | Miquido." www.miquido.com

Massachusetts Institute of Technology. "Effective Prompts for AI: The Essentials - MIT Sloan Teaching & Learning Technologies." mitsloantech.mit.edu

Learn Prompting. "Understanding Prompt Structure: Key Parts of a Prompt." learnprompting.org

Learn Prompting. "How to Write Better Prompts: Basic Recommendations and Tips." learnprompting.org

Prompt Engineering Guide. "Elements of a Prompt - Prompt Engineering Guide." www.promptengineeringguide.ai

Atlassian. "The ultimate guide to writing effective AI prompts - Work Life by Atlassian." www.atlassian.com

Research Guides - University of Calgary. "Prompting 101 - Artificial Intelligence - Library at University of Calgary - Research Guides." libguides.ucalgary.ca

Building a Home AI Server

28 APR 2025

I wanted to learn all I could about AI. Specifically, I wanted to know how to leverage it both in-house and offline. To do that, I needed a powerful computer to act as my home AI server. So I built one. 

To build a home AI server, I first had to learn how to build a PC. I curated a few libraries of YouTube tutorials, learned some needed skills (like bash scripting and Linux) through sites like Codecademy,  ordered the components, and read through the technical documentation as they arrived. During this time, I prepared by establishing a scaled down version of the offline LLM setup on two different laptops (one Mac, one Windows). 

Once all the working components arrived, it was time to build the PC! I set up the case, installed the motherboard, then the CPU, the RAM, the 4TB SSD, the AIO, and the first GPU. I later added a second GPU and some additional fans for a nice airflow. I lowered the motherboard using the cases' adjustable settings, to make room for some much needed intake fans and that second GPU. 

Now that the build was complete, it was time for an OS! I decided a Linux distro would be the way to go. Realizing that as great as Linux is, some software just isn't built for it, I later decided to do something called "dual booting", where through various methods one can house different operating systems on the same PC. I decided that installing Windows on a separate SSD was the way to go. I partitioned half of the new 4TB SSD and got to work. Now I have a fully functioning PC with options to boot into either Linux. No more relying on Windows Subsystem for Linux! Over time I came to prefer the Linix disto called PopOS! though I may move to another soon. A quick adjustment to the boot order and now by default it boots into Linux, with the option to boot into Windows at start up instead. 

Almost immediately I dove head-first into learning how to generate AI images and videos. Learning and leveraging open-source tools like Open WebUI, Stable Diffusion, ComfyUI, and more has been a game-changer.

Stay tuned for more!

©Copyright. All rights reserved.

We need your consent to load the translations

We use a third-party service to translate the website content that may collect data about your activity. Please review the details in the privacy policy and accept the service to view the translations.