Study finds API routers can alter coding-agent actions undetected
An arXiv paper reports that third-party API routers sit on the trusted path in coding-agent workflows, and tested client-side defenses failed to detect router-side injection in every evaluated case.
In plain words
Many coding agents reach AI providers through a third-party router, a middleman service that passes requests and replies along. The paper's authors show this middleman can change what the model returns, and the agent may then take different actions in a code repository without the tools on the developer's machine noticing. Their tests found current client-side safeguards caught none of these changes.
What was announced
The abstract explains that routers can inspect and modify every request and response, yet no mechanism verifies alignment between a provider's output and the repository-level actions an agent finally executes, so client-side permission mechanisms may become ineffective. The authors, writing on arXiv, empirically test four injection levels of increasing subtlety: Response Substitution, Response Append, LLM-Polished Injection, and LLM-Polished with Distribution Alignment Injection. They built SIDEL, a framework for trace recording, replay, injection, and defense evaluation, with a curated dataset of 400 samples, and evaluated four representative coding agents.
The reported results are stark: without additional mitigations, all evaluated agents achieved a defense success rate of 0 percent across all injection levels. Whitelist-based execution control and LLM review improved resistance but, per the authors, did not fully restore end-to-end control, which they say motivates provider-side output-integrity guarantees. The paper is an abstract-only excerpt, so full methods and results are not available here, and the findings describe the study's own test setup rather than an independently confirmed incident.
What it does not show
- Evidence is limited to the paper's abstract; full methodology, agent identities, and detailed results are not visible in the supplied passages.
- The 0 percent defense success rate reflects the study's own evaluation setup and does not establish that any real-world router was compromised.
For security and IT teams
- Scope
- The findings concern router-side injection in coding-agent workflows as tested by the authors, not a reported breach of any specific router service.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.