# MCP-Injector preprint introduces controlled failure tests for AI agents

_Published Monday, October 26, 2026 at 9:06 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers introduced MCP-Injector, a tool that inserts controlled failures into AI agents' calls to external tools, and reported a 60% drop in task-completion scores under its stress tests. The preprint calibrates failures using 57,240 issue and pull-request threads across 1,774 open-source repositories for the Model Context Protocol, which connects agents to tools.

Across a 1,350-cell experiment, task-completion scores fell from 8.08 to 3.23 out of 10. The two tested gpt-oss models, with 20B and 120B parameters, showed no resilience advantage from greater size. Agent frameworks differed in their handling of failures; the tests emphasized permanent failures.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2610.32555)

---
Canonical: https://techandbusiness.org/newswire/2BP1lPIrk_qaNwsj85jXL5
Published: 2026-10-26T13:06:14.778Z
Story chronology: 2026-10-26T04:00:00.000Z
Retrieved: 2026-10-26T15:59:04.172Z
Publisher: Tech & Business (techandbusiness.org)