Keras, ONNX, and GGUF model files carry hidden code execution risks beyond pickle

·1 views

A technical analysis reveals that several machine learning model file formats pose executable code risks that are widely underestimated by developers and security tooling. Keras model files store Lambda layers as marshalled Python code objects that are executed on load, making .keras and .h5 files functionally equivalent to pickle in terms of code-carrying risk. ONNX files, often considered safe due to the absence of embedded bytecode, can expose systems to arbitrary file reads via external tensor path references and may depend on non-standard operators. GGUF files embed Jinja chat templates that can be exploited through known sandbox escapes, meaning static analysis rather than rendering is required for safe inspection. The article details how the AIsbom v1.3.0 scanner addresses these risks by flagging critical issues without executing or unmarshalling any untrusted content.

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)

Silent Detectors in AI Coding Workflows May Give False Confidence, Dev Warns

A developer writing for DEV Community discovered that an AI coding agent miscounted self-tests in tooling scripts due to a case-sensitivity bug in its regex, returning a plausible but incorrect result with no indication of failure. The incident highlighted a broader problem: as developers rapidly accumulate automated detector scripts through AI agents, there is no standard mechanism to verify whether those detectors are actually functioning. Unlike traditional test suites, which can be validated through mutation testing, bespoke audit scripts and pre-commit hooks have no equivalent soundness check. The author notes that a broken detector and a working one produce identical output — silence — making it impossible to distinguish a clean result from a missed one. As a solution, the developer proposes building a self-test flag into every detector, using paired positive and negative controls to confirm the tool can both catch real issues and stay silent on false ones before reporting any verdict.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CanineIQ Tool Scores Dog Ownership Readiness to Curb Irresponsible Breed Choices

Developer Anubhav Dash built CanineIQ, a free web app that assesses whether a prospective owner is genuinely prepared for a specific dog breed, inspired by a viral CCTV video of a couple attacked by an unsupervised pitbull in India. The tool asks eight questions covering living space, experience, budget, training plans, and household safety, then generates a deterministic readiness score between 0 and 100 with a verdict of READY, CAUTION, or NOT_READY. A scoring engine weighs each answer against documented breed-specific needs, while Google's Gemini AI is used only to write a personalised narrative and a letter voiced from the dog's perspective. The app also includes a breed health transparency page highlighting genetic issues in popular dogs, such as French Bulldogs requiring palate surgery and Pugs struggling to breathe during exercise. CanineIQ was released as an open-source React application on Vercel ahead of International Dog Day, aiming to reduce breed surrenders, neglect, and bite incidents linked to underprepared owners.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Two Russian Marketplace Integrations Broke a Shared Go Interface

A development team building a self-hosted store for sellers on Ozon and Wildberries, Russia's two largest marketplaces, attempted to unify their sync logic under a single Go interface covering stock, pricing, and orders. The effort failed at every method, as the two platforms differ fundamentally in how they identify products — Ozon uses a seller-assigned offer_id, while Wildberries uses a nested nmID and barcode system requiring separate database indexes. Their APIs also diverge on response formats: Ozon returns per-item results on stock pushes, whereas Wildberries responds with a silent 204 and only flags errors in a 409 conflict. Price updates follow different patterns too, with Wildberries returning a task ID for asynchronous processing rather than an immediate result. The team concluded that a shared interface added abstraction without reducing real complexity, and published their findings alongside the open-source codebase.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Chooses Python and FastAPI to Build Collaborative Project Management Platform

A software developer has begun documenting the journey of building their third major production-grade web application, starting with the process of choosing what to build. After initially considering a live collaborative text editor, they ruled it out because its core challenges are already addressed by mature libraries and offer limited learning in backend engineering and system design. Instead, they settled on a collaborative project management platform, which naturally introduces a broad range of engineering problems such as caching, event-driven workflows, real-time synchronization, and authorization. Having previously worked with ASP.NET Core and the .NET ecosystem, the developer chose Python with FastAPI to gain exposure to a different technical ecosystem. FastAPI was selected for its built-in WebSocket support, lightweight configuration, automatic request validation, and dependency injection, making it well suited for real-time application development.

0 comments Read more at DEV Community