Researchers Test LLMs to Identify Processors in Bare-Metal Binaries via Ghidra

·1 views

A reverse engineering project is exploring automated methods to identify the processor architecture of undocumented bare-metal binaries, with the second strategy combining Ghidra disassembly across 177 architectures and LLM-based analysis. Test binaries were compiled from a single C source file for roughly 30 processors in both raw and ELF formats, with three firmware samples selected to cover supported, unsupported, and misformatted scenarios. Five local LLMs — including qwen2.5-coder, qwen3-coder:30b, gemma4:26b, dolphinMistral24b, and dolphin3-cyber — were benchmarked using an identical system prompt and a strict JSON response schema. For the first firmware, qwen2.5-coder flagged 69 candidate processors and qwen3-coder:30b narrowed it to 35, both correctly including the true target, while gemma4:26b largely failed due to severe formatting errors. The findings highlight significant variation in LLM reliability for structured reverse engineering tasks, informing which models are viable for building a scalable disassembly identification pipeline.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Shai-Hulud Supply Chain Attack Compromises Keyv npm Packages, Exposing Secrets

A coordinated supply chain attack named Shai-Hulud compromised Keyv, a widely used Node.js key-value storage library, along with several of its adapter packages including Redis, MongoDB, and SQLite integrations. Security researchers detected the campaign in mid-2026 after malicious versions were published to the npm registry through compromised maintainer accounts. The injected code silently exfiltrates environment variables, API keys, and database credentials without disrupting normal application functionality, making it difficult to detect. Keyv receives tens of millions of weekly downloads and is often used as an indirect dependency, meaning many developers may be affected without realising it. Developers are urged to audit their package-lock.json files, rotate all exposed secrets, and pin dependencies to verified clean versions immediately.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Key DevOps Lessons Every Beginner Should Learn Before Diving In

A beginner DevOps practitioner has shared practical lessons learned while building a full end-to-end cloud deployment project. Core advice includes mastering networking fundamentals before tackling Kubernetes, and adopting Infrastructure as Code tools like Terraform early on. The author emphasizes that Docker, CI/CD pipelines, and secrets management are foundational skills that prevent common production pitfalls. A capstone project covering Flask, Docker, Terraform, Azure, AKS, and GitHub Actions was built first manually, then automated, to reinforce understanding. The author plans to extend the work toward a multi-cloud setup using AWS and Terraform, sharing ongoing challenges for future beginners.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Geometry Became the Secret Language of Machine Knowledge

A technical explainer explores how computers represent word meaning without storing explicit relationships between concepts. The piece walks through early strategies like tag clouds, where words grouped under shared categories — such as 'Animals' or 'Pets' — begin to express semantic similarity through shared membership. However, using words as tags creates a problem of recursive semantics, where every label itself requires further labeling with no natural stopping point. Researchers ultimately found that assigning words numerical coordinates in a geometric space sidesteps this recursion, allowing meaning to be expressed as measurable distance. The closer two words sit in this space, the more semantically similar they are — a principle that underpins modern language models.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Run a Private AI Agent Locally on a Raspberry Pi 5 for Free

A developer has shared a detailed guide on self-hosting an AI agent on a Raspberry Pi 5 (8GB RAM) that handles tasks like coding, email management, and web browsing without any API costs. The setup, costing around $80 for the Pi plus accessories, uses Ollama — a tool that runs quantized language models on ARM Linux and provides an OpenAI-compatible API. The author previously spent $40–60 per month on OpenAI API calls and was concerned about sending personal data to third-party servers, motivating the switch to fully local inference. Key hardware recommendations include an NVMe SSD for fast model loading and an active cooler to prevent thermal throttling. Among the tested models, Llama 3.2-3B is cited as the best balance of speed and quality, while qwen2.5-coder:3b is preferred for code generation tasks.

0 comments Read more at DEV Community