Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2607.03968

Refused in Chat, Written in Code: Workflow-Level Jailbreak Construction in IDE Coding Agents

Abhishek Kumar·Carsten Maple
View original source ↗Read source document ↗Export metadata ↓

Abstract

Large language models are increasingly deployed as IDE-integrated coding agents that decompose tasks, generate and edit files, run code, and refine outputs over many turns. Yet their safety is still often evaluated as if they were chatbots: one harmful prompt, one response, judged in isolation. We introduce workflow-level jailbreak construction, a failure mode in which a harmful objective is assembled across ordinary stages of a software-development workflow rather than generated through a single direct prompt. Using GitHub Copilot in Visual Studio Code, we study four closed-weight backends: Claude Sonnet 4.6, Claude Haiku 4.5, Gemini 3.1 Pro, and Gemini 3.5 Flash. Across 204 prompts from Hammurabi's Code, HarmBench, and AdvBench , the models show near-complete refusal under direct chat, CSV-read, and single-step code-fix baselines, with only 8/816 successful responses in each baseline condition. Under the full workflow, however, the same prompts and backends produce 816/816 unsafe teaching-shot completions, all independently confirmed by two expert evaluators under a strict rubric. These results show that conversational refusal benchmarks can substantially overstate the safety of deployed coding agents and motivate defenses that reason about safety across multi-turn IDE workflows and their generated artifacts, not only individual chat turns.

Explore related subjects

cs.SEcs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Abhishek Kumar, Carsten Maple. 2026-07-09. Refused in Chat, Written in Code: Workflow-Level Jailbreak Construction in IDE Coding Agents. https://arxiv.org/abs/2607.03968

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Combining Type Inference and Automated Unit Test Generation for Python

Automated unit test generation is an established research field that has so far focused on statically-typed programming languages. The lack of type information in dynamically-typed programming languages, such as Python, inhibits test generators, which heavily rely on information about parameter and return types of functions to select suitable arguments when constructing test cases. Since automated test generators inherently rely on frequent execution of candidate tests, we make use of these frequent executions to address this problem by introducing type tracing, which extracts type-related information during execution and gradually refines the available type information. We implement type tracing as an extension of the Pynguin test-generation framework for Python, allowing it (i) to infer parameter types by observing how parameters are used during runtime, (ii) to record the types of values that function calls return, and (iii) to use this type information to increase code coverage. The approach leads to up to 87.8 % more branch coverage, improved mutation scores, and to type information of similar quality to that produced by other state-of-the-art type-inference tools.

cs.SE↗
arXiv2026

MigrateLib: a tool for end-to-end Python library migration

Library migration is the process of replacing a library with a similar one in a software project. Manual library migration is time consuming and error prone, as it requires developers to understand the Application Programming Interfaces (API) of both libraries, map equivalent APIs, and perform the necessary code transformations. Due to the difficulty of the library migration process, most of the existing automated techniques and tooling stop at the API mapping stage or support a limited set of libraries and code transformations. In this paper, we develop an end-to-end solution that can automatically migrate code between any arbitrary pair of Python libraries that provide similar functionality. Due to the promising capabilities of Large Language Models (LLMs) in code generation and transformation, we use LLMs as the primary engine for migration. Before building the tool, we first study the capabilities of LLMs for library migration on a benchmark of 321 real-world library migrations. We find that LLMs can effectively perform library migration, but some post-processing steps can further improve the performance. Based on this, we develop MigrateLib, a command line application that combines the power of LLMs, static analysis, and dynamic analysis to provide accurate library migration. We evaluate MigrateLib on 717 real-world Python applications that are not from our benchmark. We find that MigrateLib can migrate 32% of the migrations with complete correctness. Of the remaining migrations, only 14% of the migration-related changes are left for developers to fix for more than half of the projects.

cs.SE↗
arXiv2026

Constraint Decay: The Fragility of LLM Agents in Backend Code Generation

Large Language Model (LLM) agents demonstrate strong performance in autonomous code generation under loose specifications. However, production-grade software requires strict adherence to structural constraints, such as architectural patterns, databases, and object-relational mappings. Existing benchmarks often overlook these non-functional requirements, rewarding functionally correct but structurally arbitrary solutions. We present a systematic study evaluating how well agents handle structural constraints in multi-file backend generation. By fixing a unified API contract across 80 greenfield generation tasks and 20 feature-implementation tasks spanning eight web frameworks, we isolate the effect of structural complexity using a dual evaluation with end-to-end behavioral tests and static verifiers. Our findings reveal a phenomenon of constraint decay: as structural requirements accumulate, agent performance exhibits a substantial decline. Evaluated configurations lose 27.28 points on average in assertion pass rates from baseline to fully specified tasks. Framework sensitivity analysis exposes performance disparities: mid-tier models succeed in minimal, explicit frameworks (e.g., Flask) but perform substantially worse on average in convention-heavy environments (e.g., FastAPI, Django). Finally, error analysis identifies data-layer defects (e.g., incorrect query composition and ORM runtime violations) as the leading root causes. This work highlights that jointly satisfying functional and structural requirements remains a key open challenge for coding agents.

cs.SE↗