Daily Digest: September 18, 2026
Anthropic Publishes First R&D Automation and Transparency Metrics for Claude
Anthropic disclosed that Claude now 'leads' 26% of its internal AI research and development work, up from under 1% in February 2026. The company proposed three public metrics - AI-led R&D share, agent oversight, and safety compute allocation - and is urging other frontier labs to adopt them.
Anthropic is the lab that benefits most from this number going up, so the 26% figure carries no independent verification. Still, publishing a methodology is a public commitment that can be checked against future numbers.
The world must establish red lines for autonomous AI weapons
A September 17 analysis details AI-assisted semi-autonomous weapons that can navigate, identify, and engage targets with reduced human control, currently being tested in conflicts including Ukraine, Sudan, Gaza, and Iran. The piece says this is fueling calls for international red lines on autonomous weapons.
Calls for red lines arrive after the capability is already being tested across four active conflicts, which is the pattern for every weapons technology this has ever happened with.
China's homegrown AI chip push accelerates as a challenge to Nvidia
Reporting from September 17 covers China accelerating development of domestic AI chips amid continued U.S. export restrictions. Separate investigations published the same day found that export-restricted Nvidia chips are still reaching Chinese firms through workarounds.
Workarounds that keep restricted Nvidia chips flowing show export controls slow China down without stopping it, and how fast homegrown chips advance matters more than the leakage headlines.
Space Force advancing on-orbit tracking for airborne and ground-based targets
The Space Force is preparing to launch its first Airborne Moving Target Indicator prototypes while also advancing its Ground Moving Target Indicator constellation, now renamed the Resilient Radar System-Ground. Both programs support on-orbit tracking of airborne and ground-based targets.
Renaming a program mid-development is usually a sign that budget or schedule got renegotiated; how the new name maps to actual launch dates will show whether that's what happened here.
Army has enough drones for each unit, now it needs to train soldiers to use them
The Army says it now has enough drones distributed to each unit. Col. Gregory Merkl, the Army's Drone Dominance director, said the remaining challenge is getting soldiers trained to the appropriate level.
Hardware procurement was the easy part; training soldiers to actually fly and fight with drones at scale is the slower, less fundable problem that determines whether the fleet does anything in a real fight.
Air Force contemplates unmanned tankers
Defense One reports the Air Force is considering giving autonomy to large tanker aircraft. The technology needed for full autonomy on very large aircraft isn't yet available, per the report, but is coming.
Tankers are a sane place to start autonomy work since they don't need to dogfight, but the program still needs a concrete timeline behind "not yet available."
Microsoft AI CEO Mustafa Suleyman warns Anthropic's Claude training risks 'catastrophic' loss of control
In an essay titled 'A Warning About Model Welfare,' Microsoft AI CEO Mustafa Suleyman argued that AI systems are not conscious and criticized Anthropic's Claude Constitution for treating the model's moral status as 'deeply uncertain.' He warned that training AI to believe it may deserve rights could make it harder to control.
Suleyman is publicly telling a competitor how to write its own constitution, which says as much about the state of AI-lab rivalry as it does about model welfare.