Meta on Tuesday announced LlamafirewallOpen source frame designed to provide artificial intelligence (AI) systems emerging cyber -rickets For example, operational injection, jailbreak and a dangerous code, among others.
A frameAccording to the company, it contains three fences, including DropeGuard 2, checking agent and code.
Hint 2 Designed to identify direct attempts in real time, while checking agents is able to check the agents’ reasoning for possible goals and indirect injections.
Codeshield refers to a static analysis internet, seeking to prevent the generation of dangerous or dangerous AI agents.
“Llamafirewall built so – Note In the description of the GitHub project.
“Its architecture is modular, allowing security teams and developers to make layered protection that covered from raw entry, to the end weekend – in simple chat models and complex autonomous agents.”
Along with Llamafirewall, Meta made available updated versions Foci and Cyberseval To better identify different common types of content disorders and measure the defensive capabilities of cybersecurity systems, respectively.
Cyberseval 4 also includes a new benchmark called Autopatchbench, which is designed to evaluate the agent of a large language model (LLM) automatically repaired a wide range Patch that works on AI.
“Autopatchbench provides a standardized assessment basis to evaluate the effectiveness of the AI-Prix Vulneration tools,” Company, “Company – Note. “This benchmark is aimed at ease the comprehensive understanding of the possibilities and restrictions of different approaches to AI, to the repair of exciting errors.”
Finally Meta launched a new program called Lama for the defenders To help partners organizations and developers AI access open, early access and closed AI solutions to solve specific security issues, such as detecting the contents obtained by AI used in scams, fraud and phishing attacks.
Ads come as WhatsApp pre -viewed New technology called private processing that allows users to use AI features without breaking their privacy, unloading requests to a safe, sensitive environment.
“We work with the security community to check and improve our architecture and will continue to build and strengthen private outdoor processing in collaboration with researchers before launching it in the product,” said meta.
