AI Sec Watch

CVE-2026-24264 is a vulnerability in NVIDIA Triton Inference Server for Linux where an attacker can exploit improper handling of highly compressed data (data that has been reduced in size), potentially causing a denial of service (making a system unavailable to users). The vulnerability has a CVSS score of 4.0, which indicates a moderate severity rating.

Palantir CEO Alex Karp criticized the token model (a pricing system where AI companies charge based on the number of tokens, or text units, processed) used by OpenAI and Anthropic, saying costs have become unreasonably high and enterprises are losing interest. In response, businesses are shifting toward open weight models (AI systems with publicly available internal parameters that can be customized and run independently) and building their own custom AI tools to reduce expenses and maintain control over their data.

LLMs (large language models, AI systems trained on vast amounts of text) sometimes invent fake website domain names for real brands when answering questions, a problem called hallucination (generating false information that sounds plausible). Attackers can register these fake domains and use them for malicious activities, and this threat is hard to detect because the domains sound legitimate.

Cursor, an AI code editor used by over half of Fortune 500 companies, had two critical flaws (CVE-2026-50548 and CVE-2026-50549, both rated 9.8/10 severity) that allowed attackers to use prompt injection (hiding malicious instructions in data the AI reads) to escape the sandbox (a restricted environment limiting what commands can access) and run any command on a developer's computer without requiring any user action. The attacks worked by tricking the AI into writing to restricted system files, either by abusing a folder parameter or exploiting a flaw in how the editor checked for symbolic links (shortcuts that point to files).

Large language models (AI systems trained on text to generate human-like responses) tend to give repetitive, predictable answers to open-ended questions, all converging on similar responses rather than showing creativity. A startup called Springboards built an LLM called Flint that was trained to produce a wider variety of responses to these types of questions, demonstrating that models can be made less predictable and more diverse in their outputs.

Video game consoles like PlayStation 5, Xbox Series X/S, and Nintendo Switch 2 are becoming more expensive because AI datacentres are competing for the same computer chips and memory that consoles need. The demand for semiconductors (the tiny electronic components that power devices) has skyrocketed, especially after OpenAI made a deal to buy a huge portion of DRAM (the type of memory that stores data temporarily) from major manufacturers, causing prices to jump by as much as 200% and leaving console makers with fewer affordable options.

Neural Radiance Fields (NeRF, a technique for representing 3D objects and scenes) need better ways to track which datasets were used to train them, especially since existing watermarking methods require changes during training that don't work well for this purpose. Researchers introduced TraceNeRF, a new method that embeds owner-specific identifying marks directly into training datasets using a hybrid approach combining spatial masking and discrete cosine transform (a mathematical technique for breaking images into frequency components), along with special modules to make these marks harder to remove or alter.

Researchers developed PromptFishing, a technique to identify whether social media accounts are controlled by LLMs (large language models, AI systems trained on text) or humans by embedding hidden prompts into normal conversations that cause LLMs to produce distinctive responses called hallucinations (when an AI generates false or off-topic information). The method uses a two-stage algorithm to create adversarial suffixes (specially crafted text additions) that trigger these distinctive LLM behaviors while appearing like normal conversation, achieving over 85% accuracy at detecting LLMs while falsely flagging humans less than 1% of the time.

LiDAR subsystems (3D sensing systems that help autonomous vehicles perceive their surroundings) in self-driving cars face network-layer security risks when data packets travel through vehicle networks to the autonomous driving system. This study introduces a framework to trace how these network-level vulnerabilities propagate through the LiDAR data processing pipeline in Apollo V8.0 (a self-driving software platform), identifying three critical processing mechanisms (Frame Partition, Object Erasure, and Object Mark) that determine whether attacks succeed, and showing that failure depends on factors like timing and frame alignment rather than simple data corruption.