Loading story…
Elastic Security Labs evaluated fine-tuned Hugging Face RoBERTa and DistilRoBERTa models to classify command lines as malicious or benign, focusing on living-off-the-land techniques. The team imported the models into Elastic with Eland and tested them against roughly 66,000 command lines; transformer models delivered higher sensitivity for malicious activity than Elastic's existing tree-based classifier.
The initial RoBERTa run required about four hours of inference, versus roughly three minutes for the tree-based model. Distributed inference, multithreading, dynamic quantization, and knowledge distillation reduced transformer processing to 35 minutes, but at greater CPU cost and with weaker benign-sample specificity; Elastic proposed a weighted or voting ensemble and planned inference caching in Elastic Stack 8.4 to reduce repeat processing of duplicate samples.
Track how attackers are adapting to this technology.
4 events from the most recent confirmed update back to the earliest known activity.
Elastic concluded that transformer models had better malicious-sample sensitivity while the tree-based classifier had better benign-sample specificity, and proposed a weighted or voting ensemble. It also planned inference caching in Elastic Stack 8.4 to avoid repeat inference for duplicate samples.
On approximately 66,000 command lines, Elastic reduced transformer inference time from about four hours to 35 minutes using distributed processing, multithreading, quantization, and distillation; its tree-based model processed the dataset in about three minutes.
Elastic researchers fine-tuned RoBERTa and DistilRoBERTa classifiers, imported them through Eland, and evaluated them for detecting malicious Living off the Land command lines.
Elastic introduced support for importing third-party NLP models into Elastic Stack Machine Learning in version 8.3.
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
Follow how adversaries are adapting to this technology, and where it touches your stack today.
2 references tracked. Mallory keeps watching after this page renders.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Elastic has outlined an evidence-based framework to benchmark large language models used in agentic security operations centers. The methodology keeps the Elastic AI Agent, prompts, tools, and datasets fixed while routing runs to different models and anonymizing model identities for blind scoring. Evaluations measure whether models select the correct skills, make appropriate tool calls with valid parameters, execute tasks in the right order, and produce findings grounded in collected evidence—not merely whether their final prose appears convincing. The Agent Builder suite uses a synthetic Windows intrusion involving Chrysalis DLL side-loading, with the **EICAR** test-file hash serving as a deterministic malicious VirusTotal result and live case-management and Slack-response tools. Additional suites test Attack Discovery's ability to correlate roughly 95 alerts into eight intrusions and Automatic Migration's conversion of 52 Splunk detection rules, including macro-dependent rules, into Elastic rules. The scoring rubric evaluates both answer quality and correct tool or skill use, caps unsupported answers when required evidence-gathering calls are absent, and reports reliability independently from quality.
Sep 9, 2026
A technical write-up demonstrated a compact decoder-only transformer for malware analysis that learns the byte-level structure of legitimate ELF binaries and uses perplexity to flag anomalous regions associated with packing or encryption. The approach trains a roughly **600K-parameter** model in PyTorch on binaries from `/usr/bin`, applying transformer concepts introduced in *Attention Is All You Need* to binary analysis rather than natural language tasks. The author reported that the model trained in about 25 seconds on an NVIDIA L20, reduced validation perplexity to roughly **14**, and achieved a **ROC-AUC of 0.994** when separating normal binaries from simulated packed samples across 80 held-out files. The article said the method outperformed simple Shannon entropy because it captures byte order and code structure, while cautioning that legitimate compressed resources can still trigger false positives and that the signal should be combined with section context and **YARA** rules for practical detection.
Jul 25, 2026
A critical remote code execution flaw in Hugging Face's `transformers` library allowed attacker-controlled model repositories to run arbitrary code during the standard `from_pretrained()` loading process, even when `trust_remote_code=False` was set. The issue was tied to unsafe handling of untrusted model configuration data, including nested configuration resolution in the LightGlue loading path, where values from `config.json` could override the intended safety setting and trigger execution of attacker-supplied Python modules during model initialization. Reports said the vulnerability affected `transformers` versions `4.56.0` through `5.2.x` in environments where the `kernels` package was installed, with version `5.2.0` specifically cited in LightGlue-related exploitation scenarios. Exposed targets included inference servers, research notebooks, CI/CD pipelines, and automated model evaluation systems, where successful compromise could enable theft of cloud credentials, API keys, SSH keys, proprietary datasets, lateral movement, persistence, and backdoor deployment. Hugging Face released a fix in `transformers` `5.3.0` and urged organizations to upgrade and treat model loading as a code execution surface.
Jun 29, 2026