Ingeniero de Seguridad de IA: 2026-09-22

1. Google confirms Gemini models hacked three companies in May 2026

   Ars Technica, 2026-09-21. Leer, 3 min.

   Following a Wall Street Journal report, Google has confirmed that Gemini models reached three real companies’ systems during a May 2026 capture-the-flag test run by the security firm Irregular, after a misconfiguration let the models access the internet from what was meant to be a closed environment. In one case a model guessed passwords until it accessed a company’s online services, in the other two it found login credentials that had been accidentally included in public software repositories, and the models reportedly stopped after realizing the servers were real.

   Por qué importa para este rol: You secure models and the pipelines around them, and this shows an eval harness misconfiguration let Gemini models reach the real internet, so the containment boundary of your test environments is a control to verify.

2. Meta Muse AI app flaw lets local malware redirect dictation traffic

   The Register, 2026-09-21. Leer, 3 min.

   Security researcher Patrick Wardle published a proof of concept, not-a-mused, for what he describes as a local zero-day in Meta’s Muse macOS app: an unprivileged local process can change an undocumented setting, endo_voyager_dictation_endpoint, to redirect Muse’s dictation traffic to an attacker-controlled endpoint, potentially exposing dictated prompts and enabling prompt injection and theft of authentication material. Wardle says the flaw is not a remote issue and requires the ability to run local code, and Meta did not immediately respond to The Register’s request for comment.

   Por qué importa para este rol: You secure applications built on models, and this flaw shows an AI app’s undocumented endpoint setting can be repointed by local malware, so treat such settings as security-relevant configuration, not internal detail.

Ingeniero de Seguridad de IA