Right now, we’re excited to share that the SecureBERT 2.0 mannequin is out there on HuggingFace and GitHub with an accompanying analysis paper. This launch marks a major milestone, constructing on the already extensively adopted SecureBERT mannequin to unlock much more superior cybersecurity functions. Simply see this unparalleled efficiency throughout real-world duties:

In 2022, the primary SecureBERT mannequin was launched by Ehsan and a workforce of researchers from Carnegie Mellon College and UNC Charlotte as a pioneering language mannequin designed particularly for the cybersecurity area. It bridged the hole between general-purpose NLP fashions like BERT and the specialised wants of cybersecurity professionals—enabling AI methods to perceive the technical language of threats, vulnerabilities, and exploits.

By December 2023, SecureBERT ranked among the many high 100 most downloaded fashions on HuggingFace out of the roughly 500,000 fashions then out there on the repository. It gained important recognition throughout the cybersecurity group and stays in energetic use by main organizations, together with the MITRE Menace Report ATT&CK Mapper (TRAM) and CyberPeace Institute.

On this weblog, we’ll mirror on the influence of the unique SecureBERT mannequin, element the numerous developments made in SecureBERT 2.0, and discover some real-world functions of this highly effective new mannequin.

The influence of the unique SecureBERT mannequin

Safety analysts at enterprises and businesses commit an incredible period of time to parsing by way of varied safety indicators to determine, analyze, categorize, and report on potential threats. It’s an necessary course of that, when executed fully manually, is time-consuming, costly, and vulnerable to human error.

SecureBERT gave researchers and analysts a software that would course of safety experiences, malware analyses, and vulnerability write-ups with contextual accuracy by no means earlier than potential. Even at the moment, it serves as a useful software for cybersecurity specialists at a number of the world’s high businesses, universities, and labs.

Nevertheless, SecureBERT had a number of limitations. It struggled to deal with long-context inputs similar to detailed menace intelligence experiences and mixed-format information combining textual content and code. Since SecureBERT was skilled on RoBERTa-base, a basic BERT encoder with a 512-token context restrict and no FlashAttention, it was slower and extra memory-intensive throughout coaching and inference. In distinction, SecureBERT 2.0, constructed on ModernBERT, advantages from an optimized structure with prolonged context, sooner throughput, decrease latency, and decreased reminiscence utilization.

With SecureBERT 2.0, we addressed these gaps in coaching information and superior the structure to ship a mannequin that was much more succesful and contextually conscious than ever. Whereas the unique SecureBERT was a standalone base mannequin, the two.0 model contains a number of fine-tuned variants specializing in varied real-world cybersecurity functions.

Introducing SecureBERT 2.0

SecureBERT 2.0 brings larger contextual relevance and area experience for cybersecurity, understanding code sources and programming logic in a approach its predecessor merely couldn’t. The important thing here’s a coaching dataset that’s bigger, extra numerous, and strategically curated to assist the mannequin higher seize delicate safety nuances and ship extra correct, dependable, and context-aware menace evaluation.

Whereas massive autoregressive fashions similar to GPT-5 excel at producing language, encoder-based fashions like SecureBERT 2.0 are designed to know, characterize, and retrieve info with precision—a elementary want in cybersecurity. Generative fashions predict the subsequent token; encoder fashions rework complete inputs into dense, semantically wealthy embeddings that seize relationships, context, and that means with out fabricating content material.

This distinction makes SecureBERT 2.0 splendid for high-precision, security-critical functions the place factual accuracy, explainability, and velocity are paramount. Constructed on the ModernBERT structure, it makes use of hierarchical long-context encoding and multi-modal text-and-code understanding to investigate complicated menace information and supply code effectively.

Let’s check out how SecureBERT 2.0 helps safety analysts in real-world functions.

Actual world functions of SecureBERT 2.0

Think about you’re a SOC analyst tasked with investigating a suspected provide chain compromise. Historically, this is able to contain correlating open-source intelligence, inner alerts, and vulnerability experiences in a course of which might take a number of weeks of handbook information evaluation and cross-referencing.

With SecureBERT 2.0, you may merely embed all related belongings—experiences, codes, CVE information, and menace intelligence, for instance—within the system. The mannequin instantly surfaces connections between obscure indicators and beforehand unseen infrastructure patterns.

This is only one potential state of affairs of many; SecureBERT 2.0 can assist and streamline a wealth of potential safety functions:

• Menace Intelligence Correlation: Linking indicators of compromise throughout a number of sources to uncover marketing campaign patterns and adversary ways
• Incident Triage & Alert Prioritization: Embedding alerts and experiences to detect duplicates, associated incidents, or recognized CVEs—lowering noise and analyst workload
• Safe Code & Vulnerability Detection: Figuring out dangerous patterns, insecure dependencies, and potential zero-day vulnerabilities in supply code
• Semantic Search & RAG for Safety Ops: Offering context-aware retrieval throughout inner information bases, menace feeds, and documentation for sooner analyst response
• Coverage and Compliance Search: Enabling correct semantic lookup throughout massive regulatory and governance corpora

Not like generative LLMs that create textual content, SecureBERT 2.0 interprets and buildings info to ship sooner inference, decrease compute prices, and decrease the chance of hallucination. This makes it a trusted basis mannequin for enterprise, protection, and analysis environments the place precision and information integrity matter most.

Below the hood of SecureBERT 2.0

There are three elements to the SecureBERT 2.0 structure that make this mannequin such a major development: its ModernBERT basis, its information growth, and smarter method to pretraining.

SecureBERT 2.0 is powered by ModernBERT, a next-generation transformer designed for long-document processing. Prolonged consideration mechanisms and hierarchical encoding permit the mannequin to seize each fine-grained syntax and high-level construction—crucial for analyzing lengthy, multi-section safety experiences.

The mannequin is skilled on 13 occasions extra information than the unique SecureBERT with a brand new corpus that features curated safety articles and technical blogs, filtered cybersecurity information, code vulnerability repositories, and incident narratives. In complete, this dataset covers 13 billion textual content tokens and 53 million code tokens.

Lastly, a microannealing pretraining curriculum progressively transitions from curated to real-world information, balancing high quality and variety. Focused masking teaches the mannequin to foretell essential safety actions and entities like “bypass,” “encrypt,” or “CVE,” strengthening area illustration.

The efficiency of SecureBERT 2.0 is a marked enchancment over its predecessor and different evaluated fashions throughout benchmarks; the small print may be present in full analysis paper.

Trying forward: AI for safety at Cisco

SecureBERT 2.0 demonstrates what’s potential when structure and information are purpose-built for cybersecurity. It joins different fashions, just like the generative Basis-Sec-8B from Cisco’s Basis AI workforce, as a part of Cisco’s continued dedication to making use of AI responsibly inside the area of cybersecurity.

We’re excited to share this mannequin with the world, to see a number of the modern methods it will likely be embraced by the safety group, and to proceed exploring potential usages for taxonomy creation, information graph era, and different cutting-edge functions.

You may get began with the SecureBERT 2.0 mannequin on HuggingFace and GitHub at the moment, and dig into our analysis paper for extra element and efficiency benchmarking.

The way forward for cybersecurity AI is securely clever.