A use-after-free vulnerability in the Linux IPv4 TCP stack enables unauthenticated remote code execution with a 9.8 CVSS score. The flaw impacts most modern Linux distributions, posing severe risk to servers, cloud infrastructure, and IoT devices if exploited before patching.
atlasum.net
The unauthenticated remote code execution flaw exists in the Linux kernel's IPv4 TCP implementation, affecting nearly all modern Linux distributions. Security researchers warn that unpatched systems face full compromise risk if the vulnerability is exploited in the wild.
A security gap in OpenAI's authentication flow allowed attackers to hijack ChatGPT user accounts without valid credentials. The vulnerability highlights ongoing security challenges for high-traffic AI platforms handling large volumes of sensitive user data.
The two lawyers were fined after inserting hidden prompt injection commands into a court filing to manipulate AI-assisted legal research tools. The ruling is believed to be the first to impose formal penalties for malicious prompt injection use in a legal proceeding.
OpenAI will consolidate usage caps for its consumer ChatGPT and developer-focused Codex products starting in December, eliminating separate limit tracking for users who subscribe to both services. The change is part of the company's push to streamline its enterprise and developer product offerings.
The open-source Memdoor project used a multi-agent workflow to coordinate five 27-billion-parameter Qwen3 models, outperforming the larger Fable 5 model on hard coding benchmarks. The result suggests that distributed small-model teams can match or exceed the performance of larger single models for complex tasks.
Anthropic has started terminating API access for users it detects using outputs to train competing large language models, citing terms of service violations. The policy shift reflects growing friction between AI labs over the use of public API data for model training.