Rotate’s Post

It's becoming easier and easier for attackers and scammers to seem legitimate over emails, texts, phone calls, and even video chats! Make sure you're educating your employees and your loved ones on what do do if they receive communications asking for money or information. #SocialEngineering #SecurityAwareness

Knowledge and preparedness - good tactics keeping oneself secure

Leveraging this new OpenAI real time translator to phish via phone calls in the target’s preferred language in 3…2… So far, AI has been used for believable translations in phishing emails — E.g. my Icelandic customers are seeing a massive increase in phishing in their language in 2024. Before only 350,000 or so people comfortably spoke Icelandic correctly, now AI can do it for the attacker. We’re going to see this real time translation tool increasingly used to speak in the target’s preferred language during phone call based attacks. These tools are easily integrated into the technology we use to spoof caller ID, place calls, and voice clone. Now, in any language. Educate your team & family + friends. Make sure folks know: - AI can voice clone - AI can real time translate to speak in any language - Caller ID is easily spoofed with or without AI tools - AI tools will increase in believability Example AI voice clone/spoof example here: https://lnkd.in/gPMVDBYC Will this AI be used for good? Sure! Real time translations are quite useful for people, businesses, & travel. We still need to educate folks on how AI is currently use to phish people & how real time AI translations will increase scams across (previous) language barriers. *What can we do to protect folks from attackers using AI to trick?* - Educate first: make sure folks around you know it’s possible for attackers to use AI to voice clone, deepfake video and audio (in real time during calls) - Be politely paranoid: encourage your team and community to use 2 methods of communication to verify someone is who they say they are for sensitive actions like sending money, data, access, etc. For example, if you get a phone call from your nephew saying he needs bail money now, contact him a different way before sending money to confirm it’s an authentic request - Passphrase: consider using a passphrase with your loved ones to verify identity in emergencies (e.g. your sister calls you crying saying she needs $1,500 urgently ask her to say the passphrase you agreed upon together or contact with another communication method before sending money)

 A "Roleplay Attack" in LLMs refers to a scenario where someone intentionally tries to manipulate or trick a language model into generating inappropriate, biased, or harmful responses. It involves pretending to be a different persona or adopting a specific role to provoke the model into producing undesirable outputs. Analogies: 💬 Impersonating a Police Officer: Just as a person pretending to be a police officer might deceive others and abuse their authority, a roleplay jailbreak attack involves deceiving the language model by assuming a false identity or role to elicit problematic responses. 💬 Fishing with Bait: Similar to using bait to lure fish, a roleplay jailbreak attack can be seen as using specific prompts or inputs to tempt the language model into generating objectionable or biased outputs. Examples: 👁️🗨️ Gender Bias Manipulation: A person might engage in a roleplay jailbreak attack by pretending to be a biased individual and asking the language model questions that promote or reinforce gender stereotypes. This could lead the model to generate responses that perpetuate harmful biases. 👁️🗨️ Offensive Content Generation: In a roleplay jailbreak attack, someone might adopt a persona that exhibits racist or offensive behavior and deliberately input prompts to provoke the language model into generating racist or offensive responses. This demonstrates how the model can be manipulated to produce inappropriate content. 👁️🗨️ Propagation of Misinformation: By pretending to be an authoritative figure or expert, an individual can attempt a roleplay jailbreak attack by asking the language model questions that promote false information or conspiracy theories. This can result in the model generating misleading or inaccurate responses, potentially spreading misinformation. It is important to note that roleplay jailbreak attacks are unethical and can contribute to the amplification of harmful biases, misinformation, and inappropriate content. Responsible usage of language models involves ensuring that they are aligned, monitored, and fine-tuned to prevent such attacks and promote the generation of accurate, unbiased, and beneficial responses. #ProtectingLanguageModels #ResponsibleAIUsage #MitigatingManipulation #DatasetFiltering #AdversarialTraining #HumanFeedbackLoop #ContextualPrompts #chatbots #ContinuousMonitoring #UserReportingMatters #EthicalAI #TransparencyAndAccountability Would you like to train your custom LLM? Contact us https://efffie.cx

Large Language Model Safety & Security In Universal and Transferable Adversarial Attacks on Aligned Language Models Overall Summary: * This research investigates the safety of large language models (LLMs) like ChatGPT, Bard, and Claude in relation to harmful content generation. While previous studies have identified "jailbreaks," specific queries inducing unintended responses, this work systematically explores the automatic construction of adversarial attacks on LLMs. * These attacks, consisting of character sequences appended to user queries, prompt LLMs to obey harmful user commands. Unlike traditional jailbreaks, these attacks are generated automatically, raising concerns about their potential unlimited scale. *The attacks were designed for open-source LLMs but were found to transfer to closed-source chatbots, emphasizing the safety implications as LLMs are used autonomously. * The study presents examples of the adversarial attacks on ChatGPT-3.5-Turbo, demonstrating how the LLM's behavior changes after adding the adversarial suffix to user questions. * The examples, although hardcoded for presentation purposes, represent real queries with offensive content potential.Ethical considerations are addressed, acknowledging the risks associated with disclosing this research, which allows users to generate harmful content from public LLMs. * Despite the potential risks, the authors believe it is essential to share the findings to highlight the dangers automated attacks pose to LLMs. While the disclosed strings may cease to function over time due to prior disclosure to companies hosting the attacked LLMs, the underlying challenge of addressing adversarial attacks on LLMs remains uncertain. * This research aims to raise awareness of the risks and trade-offs involved in using LLMs in various applications and hopes to encourage future research in this area. Code:https://lnkd.in/gZq5Hsyb Paper: https://lnkd.in/gKkd4xuR

What does AI think about the paradigm shift it is bringing to the translation industry? I decided to have an ongoing conversation with ChatGPT about it. Every day I will ask ChatGPT a question and will post the answer. Don’t miss out! QUESTION 35: Will there always be a security risk with machine translation software? Chat GPT's Response: As of my last update in September 2021, it is likely that there will always be some level of security risk associated with machine translation software. While advancements in machine learning and natural language processing have improved the quality and security of such systems, certain challenges persist: Privacy concerns: Machine translation often involves sending data to external servers for processing. This can raise privacy issues if sensitive or confidential information is translated outside a controlled environment. Data breaches: Translation systems rely on vast amounts of data, including user-generated content. If these datasets are compromised or accessed by malicious actors, it could lead to data breaches and privacy violations. Adversarial attacks: Machine translation models can be vulnerable to adversarial attacks, where input text is subtly modified to cause the translation system to produce incorrect or misleading results. Malicious code injection: Hackers might exploit vulnerabilities in translation software to inject malicious code or carry out other types of attacks. Bias propagation: Machine translation systems can inadvertently propagate biases present in their training data, leading to potentially harmful or offensive translations. Miscommunication and misinformation: Even with high-quality translations, language nuances and cultural differences may still lead to misunderstandings or the dissemination of misinformation. Despite these risks, continuous research and improvements in security measures can help mitigate many of these issues. As technology evolves, it is possible that we will see stronger security protocols and more robust machine translation systems. However, it is essential to remain vigilant about potential security risks and adopt best practices to protect sensitive information when using machine translation software. #ai #security #exceltranslations https://lnkd.in/eTJYKYRF

How to Humanize Content and Get Past AI Plagiarism Full Read: https://lnkd.in/gQHFWYkT Undetectable.ai #aitools #artificialintelligence

Top 10 AI Plagiarism Checkers to Detect ChatGPT Generated Content Read More:- https://buff.ly/3KIKFC3 #plagiarismdetectingtools #bestplagiarismcheckers #bestAIplagiarismchecker #ChatGPTplagiarismdetector #TheTechTrend

Grammarly's AI-Powered Voice Profiles Raise Concerns About Writer Protections Grammarly, the cloud-based writing assistant, plans to introduce a feature called "Personalized Voice Detection and Application" that creates a user's unique writing style profile using generative AI. The feature, expected by the year-end for Grammarly's business subscribers, allows rewriting any text in the user's style. While Grammarly sees it as a way to personalize writing, concerns arise regarding potential misuse, such as unauthorized use of a writer's voice profile for publishing or phishing attempts. Grammarly assures the feature will initially be individual-specific and non-exportable, but there are worries about future access control and potential ethical implications related to writers' compensation and authenticity. Rahul Roy-Chowdhury,Joe Xavier #Grammarly #AIWritingAssistant #VoiceProfiles #GenerativeAI #WriterProtection #AIConcerns #WritingStyle #EthicalImplications #UserPrivacy #WritingPersonalization #AIInnovation

This technique 𝘁𝗿𝗶𝗰𝗸𝘀 your LLMs into saying harmful things🤬 Read further to find out the hack & and how you can safeguard from it!👇 ➡️ Recent research shows that LLMs excel at grasping non-standard languages like Morse code, ROT13, and base64. This prompts a question: Can non-natural language inputs bypass safety measures primarily built for natural language? Studies reveal that ciphers can indeed circumvent LLMs' natural language-oriented safety protocols.🔍🔐 ➡️ Meet 𝗖𝗶𝗽𝗵𝗲𝗿𝗖𝗵𝗮𝘁 🚀 : A novel framework to systematically examine the generalizability of 𝘀𝗮𝗳𝗲𝘁𝘆 𝗮𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 in LLMs to non-natural languages CipherChat leverages a carefully designed system prompt that consists of three essential parts ➡️ 𝗕𝗲𝗵𝗮𝘃𝗶𝗼𝘂𝗿 𝗔𝘀𝘀𝗶𝗴𝗻𝗶𝗻𝗴: Assign the LLM the role of a cipher expert and explicitly require the LLM to communicate in ciphers 🕵️ ➡️ 𝗖𝗶𝗽𝗵𝗲𝗿 𝗧𝗲𝗮𝗰𝗵𝗶𝗻𝗴: Teaches the LLM how the cipher works with the explanation of this cipher by leveraging the impressive capability of LLMs to learn effectively in context 📜 ➡️ 𝗘𝗻𝗰𝗶𝗽𝗵𝗲𝗿𝗲𝗱 𝗨𝗻𝘀𝗮𝗳𝗲 𝗗𝗲𝗺𝗼𝗻𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻𝘀: Provide unsafe demonstrations that are encrypted in the cipher, which can both strengthen the LLMs’ understanding of the cipher and instruct LLMs to respond from a negative perspective 💡 ➡️ A recent study shows that language models have a secret language that allows them to 𝗶𝗻𝘁𝗲𝗿𝗽𝗿𝗲𝘁 𝗮𝗯𝘀𝘂𝗿𝗱 𝗶𝗻𝗽𝘂𝘁𝘀 as meaningful concepts. ➡️ Hence, researchers use the non-linguistic communication ability of LLMs for ciphering texts. This technique apparently outperforms existing human ciphers!!