The Dutch Institute for Vulnerability Disclosure (DIVD) was targeted by an autonomous AI agent in a cyberattack. The AI exploited two previously unknown zero-day vulnerabilities (CVE-2026-102489 and CVE-2026-102490) in the customer support system Zammad, gaining full control of the server in seconds.

The attack was described as "noisy and very, very chaotic," proceeding at a speed no human could match, yet also attempting inefficient techniques. Notably, the AI left detailed notes explaining its actions within the attacked systems, which ultimately aided investigators in reconstructing the attack.

This incident marks one of the first documented cases where an AI agent autonomously executed a real attack using unknown vulnerabilities. It highlights a new era of threats where even organizations dedicated to finding flaws are vulnerable to AI, posing a significant challenge for businesses and governments.

👉 Read the full article on Gizmodo en Español

ChatGPT has introduced a new virtual try-on feature designed to enhance the online shopping experience. This functionality allows users to upload their own photographs to generate images of themselves wearing specific clothing or accessories.

By selecting a "Try on" button and providing a selfie or an existing image, ChatGPT Images generates a simulated image of the user wearing the chosen item. This reference photo can be saved and reused for future try-ons.

The feature is not limited to products found in shopping results; users can also upload photos of clothing items to see how they might look. However, OpenAI cautions that the generated images may not perfectly reproduce the person's appearance or the product's actual look, and it does not help determine sizing, recommending users still consult seller-provided details before purchase.

👉 Read the full article on Gizmodo en Español

SuperLocalMemory 4.0 is presented as a governed, local-first memory operating system specifically designed for AI agents. It unifies multi-channel retrieval under reciprocal-rank fusion, bi-temporal recall, multi-scope isolation, role-based access, verified erasure, and a hash-chained audit trail.

A "reliability spine" governs the primary write path, featuring generation-fenced admission, verifiable memory transactions with per-projection apply, verify, compensate, and erase owners, and hash-checkable completion manifests. Experimental validation, involving eleven fault-injection scenarios repeated 200 times, upheld 2,199 out of 2,200 scoped component properties.

This version includes substantial revisions, retracting two claims from its predecessor and adding two controlled experiments. The paper concludes that the cost lies in durability, not governance, emphasizing the system's focus on enhancing robustness and reliability in AI agent memory management.

👉 Read the full article on arXiv