Meta on Tuesday introduced LlamaFirewall, an open-source framework designed to safe synthetic intelligence (AI) programs towards rising cyber dangers akin to immediate injection, jailbreaks, and insecure code, amongst others.
The framework, the corporate mentioned, incorporates three guardrails, together with PromptGuard 2, Agent Alignment Checks, and CodeShield.
PromptGuard 2 is designed to detect direct jailbreak and immediate injection makes an attempt in real-time, whereas Agent Alignment Checks is able to inspecting agent reasoning for potential purpose hijacking and oblique immediate injection eventualities.
CodeShield refers to a web-based static evaluation engine that seeks to forestall the era of insecure or harmful code by AI brokers.
“LlamaFirewall is constructed to function a versatile, real-time guardrail framework for securing LLM-powered functions,” the corporate mentioned in a GitHub description of the undertaking.
“Its structure is modular, enabling security groups and builders to compose layered defenses that span from uncooked enter ingestion to ultimate output actions – throughout easy chat fashions and sophisticated autonomous brokers.”
Alongside LlamaFirewall, Meta has made accessible up to date variations of LlamaGuard and CyberSecEval to raised detect numerous widespread forms of violating content material and measure the defensive cybersecurity capabilities of AI programs, respectively.
CyberSecEval 4 additionally features a new benchmark referred to as AutoPatchBench, which is engineered to guage the flexibility of a giant language mannequin (LLM) agent to mechanically restore a variety of C/C++ vulnerabilities recognized by way of fuzzing, an method often known as AI-powered patching.
“AutoPatchBench offers a standardized analysis framework for assessing the effectiveness of AI-assisted vulnerability restore instruments,” the corporate mentioned. “This benchmark goals to facilitate a complete understanding of the capabilities and limitations of assorted AI-driven approaches to repairing fuzzing-found bugs.”
Lastly, Meta has launched a brand new program dubbed Llama for Defenders to assist associate organizations and AI builders entry open, early-access, and closed AI options to handle particular security challenges, akin to detecting AI-generated content material utilized in scams, fraud, and phishing assaults.
The bulletins come as WhatsApp previewed a brand new expertise referred to as Non-public Processing to permit customers to harness AI options with out compromising their privateness by offloading the requests to a safe, confidential setting.
“We’re working with the security neighborhood to audit and enhance our structure and can proceed to construct and strengthen Non-public Processing within the open, in collaboration with researchers, earlier than we launch it in product,” Meta mentioned.
