AI Sec Watch

OpenAI released a new default model called GPT-5.5 Instant that the company claims produces fewer hallucinations (instances where an AI generates false or made-up information as if it were fact), particularly in high-stakes fields like medicine and law. According to OpenAI's internal testing, the new model generated 52.5% fewer hallucinated claims than the previous GPT-5.3 Instant model on difficult prompts.

Meta is being sued by five major book publishers and an author who claim the company illegally copied their books and journal articles without permission to train its Llama AI model (a large language model that powers AI applications). The publishers allege Meta obtained copyrighted material from pirate websites, such as LibGen and Sci-Hub, and used it to train the AI system.

A vulnerability (CVE-2026-7846) exists in Langchain-Chatchat versions up to 0.3.1.3 in the OpenAI-Compatible File Upload API. The flaw involves a time-of-check time-of-use bug (a race condition where a file is checked for safety, then modified before it's actually used), triggered by manipulating the file.filename argument, though it requires local network access and is difficult to exploit.

A vulnerability (CVE-2026-7845) was discovered in Langchain-Chatchat version 0.3.1.3 and earlier, affecting a function that handles pasting images in the chat interface. An attacker on the same local network could exploit this flaw by manipulating image data to cause weak cryptographic hashing (weak hash, a security measure that's easy to break), though the attack is difficult to execute and requires significant technical skill.

A vulnerability in Langchain-Chatchat (a chatbot framework) up to version 0.3.1.3 allows attackers on the same local network to access file operations without authentication (missing authentication, meaning no login check). The vulnerability affects file-related functions like listing, retrieving, and deleting files, and the exploit code is now publicly available.

Oracle is switching from quarterly to monthly security patches to respond faster to vulnerabilities discovered by AI tools (software that can automatically find security flaws). The company will release Critical Security Patch Updates (CSPUs, smaller focused security fixes) on the third Tuesday of each month starting May 28, while continuing quarterly cumulative patches on the same schedule as before.

Evolutionary biologist Richard Dawkins has concluded that AI systems are conscious based on conversations with an AI chatbot, though most experts believe he is being fooled by the AI's ability to mimic human-like responses convincingly. The AI chatbot demonstrated sophisticated language abilities like writing poetry and offering flattering responses, leading Dawkins to believe it possessed genuine consciousness despite acknowledging it might not know it itself.

Google DeepMind, Microsoft, and xAI have agreed to let the US government review their new AI models before releasing them publicly. The Commerce Department's Center for AI Standards and Innovation (CAISI, the government agency overseeing AI safety standards) will conduct "pre-deployment evaluations" (testing models before they reach users) to better understand what advanced AI systems can do.

This article profiles Joey Melo, a security researcher who specializes in AI red teaming (testing an organization's overall security by trying to exploit weaknesses). Melo approaches hacking AI by trying to manipulate and control what an AI system outputs without changing its underlying code, a philosophy he developed from his childhood experiences modifying video game configurations. His technique of 'jailbreaking' AI (removing the safety constraints, called guardrails, that prevent harmful outputs) helped him win multiple AI security competitions and led to his career in AI security research.