Agentic Security @AgenticScanner

Exciting news! Agentic Security, our open-source LLM vulnerability scanner and AI red teaming kit, is gearing up for launch! Designed to safeguard AI systems from jailbreaks, fuzzing, and multimodal attacks, Agentic Security is committed to enhancing AI safety and reliability.

@shawmakesmagic 🖐️

AI is changing cybersecurity — so enterprises must rethink how they protect data and users. In the latest AI + a16z Podcast, a16z’s Joel de la Garza speaks with cybersecurity leaders about the risks of DeepSeek, deepfake threats, and what the future of security looks like. 👇

@shawmakesmagic @shawmakesmagic we might be able to help trace vulnerabilities.

Addressing the rising security challenges in AI agents and Large Language Models (LLMs) is paramount. By implementing robust defense strategies, conducting regular security audits, and fostering industry-wide collaboration, we can mitigate vulnerabilities and ensure safer AI deployments. #Cybersecurity #InfoSec #AI #CyberDefense #DataProtection #Malware #CyberThreats #NetworkSecurity #EthicalHacking

Our goal? Build safer, more secure AI systems by proactively identifying and fixing weaknesses. Join us in shaping the future of AI security! #CyberSec #LLMSecurity #SecureAI #TechTrends #AIEthics #CyberDefense

⭐ ⭐ ⭐ Day 12--Securing #AgenticAI Learning Alert ⭐ ⭐ ⭐ Working with my colleague David Linthicum, we’re going to review the OWASP Top 10 critical vulnerabilities for LLMs and provide practical suggestions for mitigation of these vulnerabilities Today's Topic: Model Theft in Agentic AI 📌 What is model theft in Agentic AI? Model theft in the context of agentic AI refers to the unauthorized act of stealing or copying a trained machine learning model, including its parameters and architecture, from another entity, essentially allowing the thief to replicate the model's capabilities without having to go through the process of training it themselves, often with the goal of gaining a competitive advantage or exploiting sensitive information embedded within the model. Stealing a model is considered a form of intellectual property theft, as the model represents significant development effort and valuable knowledge. How can attackers attempt to steal a model and what is the impact? Attackers can attempt to steal a model 3 different ways: 1️⃣ Querying the model repeatedly-- By feeding carefully crafted inputs and analyzing the outputs, attackers can infer aspects of the model's internal workings. 2️⃣ Reverse engineering--Disassembling the model's code to extract its parameters and architecture. 3️⃣ Side-channel attacks-- Exploiting unintended information leaks like electromagnetic signals emitted by the system running the model. There are two main impacts of model theft in agentic AI: ✔️ Competitive disadvantage--A competitor can gain access to a highly developed model, potentially allowing them to quickly catch up in the market. ✔️ Data privacy concerns--If the model was trained on sensitive data, stealing it could expose private information. What are the ways to mitigate model theft in Agentic AI? Today there are 4 ways to mitigate model theft in Agentic AI 🗝️ Model obfuscation--Encrypting or otherwise disguising the model parameters to make them harder to reverse engineer. 🗝️ Access control--Limiting who can access the model and how they can interact with it. 🗝️ Monitoring for suspicious activity--Detecting patterns of queries that could indicate an attempt to extract model information. 🗝️ Watermarking--Embedding unique identifiers within the model to trace its origin if stolen. #AI #AISecurity #CyberAI

Smart AI is great. Secure AI is better. An AI that retrieves and refines insights is powerful—but what happens when it retrieves sensitive data, hallucinates, or gets manipulated? At Agentic Security, we ensure AI not only thinks fast but thinks safe. Test your LLM’s security today. 🔗 github.com/msoedov/agenti…

Data cleaning isn’t just about accuracy—it’s about security too. Messy data doesn’t just lead to bad predictions; it can introduce vulnerabilities like bias exploitation, data poisoning, and unintended sensitive info leaks. At Agentic Security, we help ensure your AI models aren’t just accurate, but also safe from adversarial threats. Want to test your LLM’s resilience? Scan it today! github.com/msoedov/agenti…

Sensitive Information Disclosure is one of the most overlooked yet high-impact risks in Agentic AI. At Agentic Security, we don’t just talk about these risks—we help you detect and mitigate them. Our LLM vulnerability scanner identifies potential leaks before they become compliance nightmares. Want to test if your AI is leaking sensitive data? Try Agentic Security and scan your LLM today! github.com/msoedov/agenti… #AI #CyberSecurity #LLMSecurity #DataPrivacy

@johnrush We'd love to audit your AI agents. Contact us for a free vulnerability scan !

As AI systems become more integrated into our daily lives, ensuring their security is paramount. Recent studies have highlighted vulnerabilities in Large Language Models (LLMs), such as prompt injection attacks and data leaks. It's crucial to implement robust security measures to protect sensitive information and maintain trust in AI technologies. #AI #CyberSecurity #LLMSecurity #DataProtection

For those who haven't come across it yet, here's a handy trick to discuss an entire GitHub repo with an LLM: => Just replace "github" with "gitingest" in the url, and you get the whole repo as a single string that you can then paste in your LLMs

In today's digital landscape, robust #CyberSecurity is essential to protect against evolving threats. Embracing #AI and #MachineLearning enhances our defense strategies, ensuring a safer online environment for all. Stay vigilant and proactive. #InfoSec #DataProtection #TechInnovation

Think your AI is unbreakable? Meet Agentic Security: the open-source LLM vulnerability scanner that's ready to challenge that notion. From multi-modal attacks to multi-step jailbreaks, it's time to put your AI's defenses to the ultimate test. Are you prepared? #AI #LLM

Good security isn’t just about stopping hackers—it’s about stopping mistakes, exploits, and unintended chaos. If your system trusts anything blindly, it’s only a matter of time before it gets burned. Test, patch, and always assume the worst.

@KevinHenrikson We make sure that the LLM is secure... @AgenticScanner

Important thread on AI

Dan Hendrycks @hendrycks

2 years ago

We’ve found as AIs get smarter, they develop their own coherent value systems. For example they value lives in Pakistan > India > China > US These are not just random biases, but internally consistent values that shape their behavior, with many implications for AI alignment. 🧵

keyboard_arrow_left Previous keyboard_arrow_right Next

689 2K 11K 6.2M 10K

Some things in life should NEVER be allowed. Like running : import os; os.system("rm -rf /"). One wrong move, and BOOM—your entire system is gone. If your code allows this, you're not writing software, you're writing a disaster plan. #AI #Security

@lostgirldev @soleng_agent let's collab ? #AI #Security

Agentic Security @AgenticScanner

2 years ago

7 Critical Vulnerabilities in AI Models Prompt Injection Attacks Attackers manipulate AI by crafting inputs that override its instructions. Impact: Can make AI reveal secrets & leak data Fix: Implement strict prompt filtering & user input validation.

1 0 0 32 0