Meta announced a series of New security tools For artificial intelligence models, with the aim of doing the development and use of safer, both for creators and defenders in the field of Cybersecurity.
The news is specifically related to Llama Family model, which is now updated and more sophisticated resources to deal with emerging threats.
Llama Guard 4: Multimodal Security for text and images in Meta's new AI program
One of the major updates is represented by Llama Guard 4The evolution of the customizable security filter of the meta.
The great change of this version is its multimodal ability, which means the ability to study and apply security policies not only to the text but also the images. This step is important, considering that AI applications are becoming more visual.
Llama Guard 4 has already been included in the new Api llama By meta, available current in limited previews. Developers can access this tool through the official Llama protection page, or by Hugging face and Github Platforms.
Another significant change is the llamafirewall, a system designed to serve as Command Center of security in artificial intelligence systems. This tool allows for coordination of various protection models and is integrated with other security tools from meta.
Llamafirewall is designed to counter sophisticated threats such as Prompt injectionthe generation of potentially dangerous code, and dangerous AI plug-in behavior.
In essence, it represents a bulwark against the most unbelievable attack that can compromise the integrity of systems based on artificial intelligence.
Meta also updated its system for discovery Jailbreak Attempts and immediate injection in the new Prompt Guard 2 (86m). This model is designed to more accurately identify attempts to manipulate AI.
Next to it, PROMPT GUARD 2 22M is introduced, a lighter and faster version. With a reduction in latency and computing costs up to 75%This version is ideal for those who work with limited resources but do not want to compromise security.
“` Html
Cybersec Eval 4: New Benchmarks for AI security
“`
Meta not only provided tools for developers, but also updated the Cybersec Eval 4 Benchmark Suite, which was designed to assess the capabilities of AI systems in the field of cybersecurity.
The open resource suite helps organizations measure the effectiveness of artificial intelligence in removing and responding to digital threats.
Two new tools enriched this suite:
- – Cybersoc Review: Formed in collaboration with CrowdstrikeThis framework evaluates AI's performance in a true security operation center (SOC) context, which offers a concrete view of the effectiveness of running artificial intelligence.
- – Autopatchbench: a benchmark that tests the ability of AI models, including those from the Llama family, up to Automatically identify and correct the weaknesses In the code before they are being exploited by malicious actors.
To facilitate the adoption of these tools, Meta launched the Llama Defenders Program, which offers privilege access to a selection of AI solutions -some open resources, others in preview or ownership -designed to deal with specific challenges in the field of security.
Among the shared tools are also the Automatic classification of sensitive documentsused inside the meta.
This system has applied security labels to documents within an organization, which prevents confidential information from accidental entry into AI systems where it can be exposed.
Meta also addressed the growing problem of fake audio generated by artificial intelligenceespecially used in scams and phishing attempts. Two new tools are made available to partners:
- – Llama formed the audio detector
- – Llama audio watermark detector
These tools help to recognize synthetic voices in weak -call calls. Companies like Zendesk, Bell Canadaand And & t is already evaluated the integration of these technologies into their security systems.
Private process
Finally, Meta gave a preview of a technology under development for WhatsappCalled private processing.
The goal is to enable artificial intelligence to provide useful features – such as summarizing unread messages or suggestions of responses – without either meta or WhatsApp to access the content of the messages.
This technology represents an important step toward a The respect of privacy ai. Meta adopted a transparent approach, which publishes itself Model threat and inviting the research community to test its stability before the official release.
In this series of announcements, Meta presents a concrete commitment to strengthening the security of artificial intelligence, both from views of development and defense.
The goal is twice. Especially, to protect finishing users and provide developers and security professionals with advanced tools to deal with ongoing emerging digital threats.
In a rapid change of technological scenery, where AI plays an increasingly key role, initiatives like this are important to ensure a safer, more pronounced, and responsible digital future.
