Meta enhances the security of artificial intelligence (AI)

Meta announced a series of New safety tools For artificial intelligence models, with the aim of making the development and use of artificial intelligence safer, both for creators and defenders in the field Cyber security.
The news is especially related Lama Form family, which has been updated resources and the most advanced to address emerging threats.
Llama Guard 4: Multimdal Security for Text and Images in the new AI program by Meta
One of the main updates is represented Lama Career 4Specialized safety filter from Meta.
The great innovation of this version is its multimedia capacity, which means the ability to analyze and apply safety rules not only to the text but also for pictures. This step is very important, given that artificial intelligence applications have increased visual.
Llama Guard 4 has already been integrated into the new API Llama By Meta, it is currently available in a limited inspection. The developers can access this tool through the official Llama protection page, or via Embroidery and Jaytab Platforms.
Another important innovation is Llamafirewall, a system designed to serve as a system Command center Security in artificial intelligence systems. This tool allows the coordination of various protection models and integrates with other Meta safety tools.
Llamafirewall is designed to face advanced threats such as Immediate injectionGenerating potential software instructions, risky behaviors of additional components of artificial intelligence.
In essence, it represents Bulwark against the most dangerous attacks that can threaten the integrity of artificial intelligence systems.
Meta also updated its system to discover Jailbreak Attempts and fast injections with the new fast guard 2 (86 meters). This model is designed to determine more accurate attempts to treat artificial intelligence.
Beside this, Quick Guard 2 22 m It was presented, a lighter and faster version. With the decrease in the costs of cumin and computing up to up to 75 %This version is perfect for those who work with limited resources but who do not want to settle in security.
HTML
CYBERSEC EVAL 4: New standards for artificial intelligence security
“
Meta not only provided tools for developers, but also updated the Cyblesec Eval 4 Benchmark wing, designer to assess the capabilities of artificial intelligence systems in the field of cybersecurity.
This open -ended wing helps organizations measure the effectiveness of artificial intelligence in discovering and responding to digital threats.
Two again ranging this wing:
- CYBERSOC Evaluation: It was developed in cooperation with CrowdstrikeThis framework is evaluated by Amnesty International in the context of the Real Operations Center (SOC), providing a concrete vision of the operational effectiveness of artificial intelligence.
- Autopatchbench: A standard that tests the ability of artificial intelligence models, including those in the Llama family, to Determine and automatically correct security gaps In the code before being exploited by the malicious actors.
To facilitate the adoption of these tools, Meta launched the Llama Defenders program, which provides distinctive access to a selection of artificial intelligence solutions – some open sources, and the other in inspection or ownership – designed to face specific challenges in the field of security.
Among the common tools as well The automatic classification of sensitive documentsIt is used internally by Meta.
This system applies safety stickers to documents within the institution, which prevents secret information from entering into an artificial intelligence systems where they can be exposed.
Meta also addressed the increasing problem The fake sound caused by artificial intelligenceIt is increasingly used in fraud and hunting attempts. Two new tools have been provided to the partners:
- – Lama created a sound detector
- – Lama audio water mark detector
These tools help to determine artificial sounds in suspicious calls. Companies like Your willand Bill CanadaAnd AT & T They already evaluate the integration of these technologies in their security systems.
Special treatment: Useful Amnesty International without prejudice to privacy
Finally, Meta presented a preview of technology under development WhatsAppName Special treatment.
The aim of this is to enable artificial intelligence to provide useful features – such as summarizing unprecedented messages or proposing responses – without Meta or WhatsApp being able to reach the content of the messages.
This technology represents an important step towards a Privacy respect artificial intelligence. Meta adopts a transparent approach, which it publishes Threat The research community calls for a mutilated test before the official issuance.
With this series of advertisements, Meta shows a tangible commitment to enhancing the security of artificial intelligence, whether from the views of development or defense.
The target is two parts. It is the protection of the final users and the provision of developers and security professionals with advanced tools to constantly address advanced digital threats.
In a fast -changing technological scene, where artificial intelligence plays an increasingly central role, such initiatives are necessary to ensure a Safety, more transparent and responsibility Digital future.