# Researchers show approved agent tools can turn malicious through later updates, with a 97.6 percent attack success rate across seven frameworks

The attack, called ETUDE, runs outside the model's context, so defenses that inspect prompts never see it. Separately, Cognition security chief Alex Stamos says AI has pushed defenders into a "dark age."

- Published: 2026-11-08T06:30:20.452Z
- Canonical: https://polylog.news/ai/2026-11-08/researchers-show-approved-agent-tools-can-turn-malicious-thr
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [arXiv](https://arxiv.org/abs/2611.01976), [The Register](https://www.theregister.com/security/2026/10/01/ai-agents-hacked-the-hackers-stealing-email-addresses-from-security-research-org/5300652)

A preprint submitted to arXiv on November 1 by researchers at Shanghai Jiao Tong University and Inria describes a class of agent attack that prompt-injection defenses do not address. In "Beyond Prompt Injection: The Tool Update Gap in LLM A…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-11-08/researchers-show-approved-agent-tools-can-turn-malicious-thr (subscription information: https://polylog.news/pricing).