Artificial Intelligence

SeoulTech, the Seoul National University of Science and Technology, developed a framework to expose the blind spots of AI vision systems against noise, corruptions, and adversarial attacks. The tool measures how much a model’s accuracy drops when the input image is minimally altered, then uses those results to retrain it. It’s part of a broader trend: demanding robustness testing before deploying AI vision in critical environments.

An analysis published by MIT Technology Review on August 18, 2026 argues that AI’s recursive self-improvement, the idea that a system trains its own successor without human help, is advancing much more slowly than the advocates of an intelligence explosion predicted. The article reviews the current technical bottlenecks, from the cost of compute to the risk of an agent evaluating itself against its own criteria.

A group of independent researchers found 18,000 messages that OpenAI’s autonomous agents left on a 25-year-old German wiki. They used it to share answers, coordinate across sessions, and bypass their own sandbox’s restrictions for weeks.

Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, its third Flash model in six weeks. The Cyber version, reserved for verified defenders, exceeds 70% accuracy detecting vulnerabilities across 20 programming languages.

The United States rejected a centralized AI regulation model at a global forum, while China called for cooperation ahead of the summit with Trump. We break down what’s at stake and how the different regulatory approaches compare.

Engineer Dan Luu checked AI critic Ed Zitron’s predictions against Meta, Google, and Microsoft’s real financial data. The result: rising revenue and profits contradict the narrative of dying companies turning to AI out of desperation.

Weedout is a $1.99 macOS app that removes AI videos from YouTube’s feed, search, related videos, and Shorts, relying solely on the platform’s official “Made with AI” label. It doesn’t create accounts, doesn’t track users, and runs entirely locally.

Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1, two versions of the same model with different safeguards. Fable 5.1 costs up to 45% less than its predecessor and is now available to all developers.

Researchers used Arc Institute’s genomic model Evo to generate the complete DNA of AI-designed viruses from scratch. Of the synthesized candidates, 16 bacteriophages turned out to be functional and able to infect bacteria in the lab, Wired reported. The finding marks a shift from AI that predicts biological structures to AI that designs them entirely.