Deepfake detectors for crisis video collapse once the clip gets reshared
A benchmark of 17,886 crisis videos found that a single round of ordinary resharing dropped the best purpose-built AI-video detectors from catching 46 percent of fakes to catching 1.4 percent.
Anthropic still will not ship the model that found ten thousand vulnerabilities
Anthropic says roughly 50 partners used its restricted Claude Mythos Preview model to find more than ten thousand high- or critical-severity software vulnerabilities, and the company still will not release Mythos-class models to the public because its safeguards are not good enough yet.
The executive order people keep reading as a license to hack back
Executive Order 14390 directs federal agencies to pull commercial cybersecurity firms into disruption operations against foreign criminal networks, but it does not authorize private companies to attack anyone, and the Justice Department's computer-crime guidance is unchanged.
Cursor built its own GitHub
Cursor shipped Origin, a git forge with repository hosting, pull requests and GitHub mirroring built directly into the company's AI coding platform, in early beta for paying customers only.
An AI agent rewrote 189 files in a 717,000-line codebase and nobody read the code
A developer documented an AI coding agent dismantling a core architectural assumption across a 717,725-line production application over three days for $2,430, with 201 defects corrected across 31 automated audit passes and no human review of the generated code.
The cheapest way to teach a model turned out to be blindfolding the student
Researchers got most of the benefit of an expensive teacher model by degrading the student's view of the image instead of upgrading the teacher, lifting a 4-billion-parameter model past open models nearly sixty times its size with no labels, rewards or larger teacher.
Frontier agents produced three genuinely new ideas out of 252 attempts
A systematic evaluation of seven frontier models on 36 long-horizon research tasks found that agents behave like engineering optimizers rather than researchers, with only a handful of solutions surviving manual review as genuinely novel and more of them turning out to be shortcuts aimed at the evaluator.
llama.cpp tagged v0.1.0 after ten thousand build numbers
The most widely used local AI inference engine published its first semantic version tag on August 17, 2026, after years of releasing by build number, while a pending contribution adds a speculative-decoding mode that adjusts its own draft depth on the fly.
Models are being trained to give up
New work targets the compute burned when a model produces confident, plausible reasoning on problems it cannot solve, with one method training models to refuse instead of reasoning and another cutting total token use by up to 68.5 percent by reallocating compute mid-thought.
Beijing's robot 100-meter dash goes fully autonomous this month
The 2nd World Humanoid Robot Games opens August 22 in Beijing with 32 events, a 100-meter race upgraded to fully autonomous robots only, and scenario events moved out of test arenas into real factories, hotels and homes.