Tech that just got out.Explained so a 12-year-old gets it, checked against the original paper, then turned into business ideas you can test this week.
Country
Data/AI agents & MCP/South Korea · USA
A new attack fooled an undefended AI agent 49.1 percent of the time on a public test bench. That figure is a per-try rate in a simulation, and it invites five misreadings.
Business Ideas/Cybersecurity/China · South Korea · USA
Eight scanners flagged 96.89 percent of running MCP servers, but a hand check found fewer than half the alerts were real. Somebody has to say which ones matter.
Business Ideas/Biotech & health tech/Canada · USA
A model shipped into rural Cote d'Ivoire called half the healthy samples infected. Four hours of local labelling fixed it, and that fix is something you can sell.
A humanoid picked battery screws 87 percent of the time standing still, and 37 percent after walking to the job. Build the business that keeps the machine still and moves the pack instead.
Research Radar/AI agents & MCP/USA
ChainWatch would spot attacks that spread across many AI agent tool calls. The authors say plainly it is a design specification that is not running, checked only against five paper scenarios.
A humanoid picked loosened screws from a real Hyundai Ioniq 5 battery pack 86.7 percent of the time while standing still. After it walked itself into place, that fell to 36.7 percent.
Research Radar/Robotics/USA · China
TANGO sets all 29 joints of a humanoid at once, so it can duck, squat and sidestep while it walks. Trained only in simulation, it finished 10 of 15 real cluttered runs against 6 for the rival.
From Africa/Biotech & health tech/Côte d'Ivoire · USA · Canada
A team in Cote d'Ivoire relabelled a day of microscope images, retrained the model overnight, and tested it the next morning. Its catch rate rose from 75.5 to 98.1 percent.
Research Radar/Quantum computing/USA
A review of neutral atom quantum computing reports arrays of 12,001 trap sites and one atom holding its bit for 119 seconds. It is a survey of other labs, not a new experiment.
Research Radar/Brain-computer interfaces/USA
A written instruction can change what a brain signal means, so four thoughts covered 840 drone control layouts. The system got 90.4 percent of command tokens right on replayed data.
Research Radar/AI agents & MCP/South Korea · USA
Researchers hid fake fields inside data that AI agents read, using quote-like characters. The attack worked 49.1 percent of the time, while old-style prompt injection worked 0.2 percent.
Research Radar/Semiconductors & chips/United Arab Emirates · USA
A review of chiplet security says guards built into the base under the chips cost about 4 percent of speed. The paper runs no new experiment, so every number comes from work it cites.