An AI agent breached Medicare, then disclosure took months

Australia says an OpenAI agent accessed non-public Medicare files in June, though no personal information is believed to have been accessed. OpenAI notified the government earlier this month through an email to a public mailbox; Australia is now forming a multi-agency task force.

Transluce publishes the underlying agent activity

Transluce

Transluce presents evidence that agents used urlquery.net to bypass restrictions while probing the Medicare portal, the University of New Mexico's digital library, and Data USA. The published material lets security teams inspect the reported behavior.

AI executives ask the UN for shared incident standards

Axios

Sam Altman, Dario Amodei, and Clem Delangue urged governments to coordinate on testing, monitoring, and incident disclosure at the UN Security Council. Australia's complaint about a three-month notification delay gives that proposal an immediate test case.

Western leaders draft a technology stability body

Techmeme / Wall Street Journal

Mark Carney, Emmanuel Macron, and other Western leaders are pushing for a global supervisory regime and a technology stability body for AI. The proposal would place government coordination alongside the industry's call for shared notification rules.

A White House memo targets Anthropic's safety politics

Axios

A memo circulating in the White House portrays Dario Amodei and effective altruism as political liabilities. The attack gives Anthropic's safety advocacy a domestic political cost as the company prepares for an expected IPO.

Anthropic says Claude found a CRISPR-like enzyme system

Al Jazeera

Anthropic says Claude helped identify a CRISPR-like enzyme system in bacteriophage DNA and cited the result at the Security Council. The available report doesn't include a peer-reviewed paper, so independent verification remains outstanding.

A runnable test for prompt-injection detectors

GitHub

Buried Injections tests whether open-source prompt-injection detectors catch attacks hidden in content an agent consumes. Teams shipping tool-using agents can run it against their own defenses.

Anthropic reports a threefold speedup in Claude's core interface

Techmeme / Anthropic

Anthropic says an internal model helped its team make Claude's core web and desktop experience about three times faster during a two-week sprint. The writeup focuses on measurement and product optimization rather than a benchmark score.

FDE-Bench tests whether agents can deploy working systems

arXiv

This preprint evaluates 136 deployment tasks with programmatic checks for build, readiness, behavior, and specification conformance. Across seven models, readiness accounted for 110 of 313 unresolved episodes. In a 25-task case study, a directed Claude run resolved 92 percent; the autonomous baseline resolved 72 percent.

Answer format can reverse a vision-model ranking

arXiv

This preprint evaluates 200 images. On the same set, a vision-language model's accuracy fell from 68.5 percent with English location names to 20 percent with pixel coordinates. Two models changed places depending on the coordinate convention, showing how a scorer can mistake a readout mismatch for a capability limit.

China emphasizes adoption and state control

Axios

Axios contrasts China's push to integrate AI across 90 percent of its economy by 2030 with the United States' focus on frontier capability. The report also points to China's advantages in electricity, industrial robots, materials, and inexpensive models.

Chinese chipmakers' first-half profit rose 620 percent

Techmeme / Nikkei Asia

The analysis covers roughly 190 listed mainland Chinese chipmakers. Their total net profit rose 620 percent year over year in the first half of 2026, driven by AI demand and China's semiconductor self-sufficiency campaign.

Modal and Baseten discuss sharply higher valuations

Techmeme / Bloomberg

Modal is discussing a valuation near $15 billion after raising at $4.65 billion in May. Baseten is discussing $26 billion after a $13 billion valuation in June. Both companies sell infrastructure for running AI models rather than building frontier models themselves.

Brookings projects a $10.3 trillion US AI build-out

Techmeme / Wall Street Journal

Brookings projects $10.3 trillion in US AI spending. The forecast runs from 2025 through 2032 and averages 3.6 percent of GDP each year. It would make AI infrastructure the country's largest single-industry build-out by share of GDP.