Legacy Code

Security researchers warn that advanced AI models capable of identifying and exploiting computer system vulnerabilities are becoming increasingly common and will soon be the norm across the industry. These capabilities emerge naturally from large-scale language models trained on technical data, and no regulatory mechanism can prevent their development.

Anthropic has published research on agentic coding—AI systems that autonomously write, test, and refactor code with minimal human supervision. The research demonstrates that maintaining expert-level code quality while scaling to autonomous operation requires careful attention to how AI systems approach coding problems and verify their own work.

The Trump administration has forced Anthropic to take its newest AI models, Fable 5 and Mythos 5, offline after ordering the company to block access for all foreign nationals—including international users and the company's own foreign employees. The order stems from unclear export control regulations that caught Anthropic by surprise and forced an immediate shutdown of services.

After months of aggressive AI adoption and spending, enterprises are discovering that actual return on investment is much harder to achieve than anticipated. According to NEA venture investor Tiffany Luck, companies that pushed employees to maximize AI usage through "tokenmaxxing" campaigns are now facing massive bills and cutting back on licenses.

Google has released a new smart speaker powered by its Gemini AI assistant, marking the company's first major hardware update in six years. The device was redesigned from the ground up to serve as a hub for Gemini's conversational capabilities, positioning it directly against Apple's HomePod and Amazon's Echo lineup.

IO-AI Tech, operating in Shenzhen, is training workers to control humanoid robots using VR interfaces that simulate the physical experience of operating the robots remotely. Workers don VR rigs and control the humanoid's movements in real-time, creating a hands-on workforce trained in robot operation.

OpenAI and Molecule.one have demonstrated a near-autonomous AI chemist powered by GPT-5.4 that successfully improved a challenging reaction used in medicinal chemistry research. The system operated with minimal human intervention to optimize a key drug-making process, showing how frontier AI models can handle complex real-world scientific work.

OpenAI has released LifeSciBench, an expert-authored and expert-reviewed benchmark designed to evaluate how AI systems perform on actual life science research tasks and decision-making challenges. Unlike generic benchmarks, this tool measures AI performance on problems that matter to working scientists and researchers.

A new Pew Research poll shows that while 49 percent of Americans use AI chatbots regularly, 63 percent believe the technology is advancing too quickly for society to safely manage. This concern has grown significantly since 2024, when only 33 percent of Americans reported using chatbots at all.

Trump administration officials have told Wired that Anthropic must guarantee its Fable 5 model cannot be jailbroken or have its safeguards circumvented before the company can release the system. Security experts, however, say this requirement is essentially impossible to meet—no AI system can guarantee absolute immunity to every conceivable attack.