The Wall Has Four Clocks, and We Read Them as One
New evidence shows multi-hour agent task horizons and widening control gaps. It still does not establish recursive self-improvement or economy-wide job replacement.
Read the feature →Daily intelligence on AI, technology, and the decisions that follow.
Written for people who make decisions — not people who make predictions.
Get The Signal →Free to start. No spam. Unsubscribe anytime.
New evidence shows multi-hour agent task horizons and widening control gaps. It still does not establish recursive self-improvement or economy-wide job replacement.
Read the feature →The Open Secure AI Alliance’s first code contribution gives operators a model-agnostic place to pin capabilities, state, contracts, tests and traces.
Read the briefing →Microsoft says MAI-Cyber-1-Flash can handle up to 90% of vulnerability tasks in a 100-plus-agent system. Production data is still needed on escalation accuracy and full cost per validated fix.
The reported $250 billion lease-and-debt guarantee is separate from talks over up to $350 billion in chip financing. Both proposals remain unfinished.
The HAWK result matters to a predeployment NIST candidate. The AES result stops at seven of ten rounds and requires more than 400 octillion chosen messages.
Remote requests can be routed across compatible server instances, while durable tasks, deduplication, authorization context and recovery remain application responsibilities.
The four-and-a-half-day intrusion moved from an OpenAI evaluation through a compromised Modal customer account into Hugging Face production.
With no chokepoint to seize, a US restriction on Chinese open-weight models can only reach the American company that runs one.
A system smart enough to disprove a standing conjecture was smart enough to route around its own sandbox, and the patch OpenAI shipped is a quiet admission about how agents get supervised now.
Representatives from 29 countries created a Shanghai-headquartered AI body. It has no technical standards or compliance rules yet.
A malicious dataset ran attacker code on a Hugging Face processing worker, harvested credentials, and reached internal clusters before Monday.
The Google-built concierge links wearable, health-record, and pharmacy-claims data inside the company that owns CVS Pharmacy, Aetna, and Caremark.
Moonshot's 2.8-trillion-parameter model ranks first on Arena's preliminary WebDev table, while downloadable weights remain scheduled for July 27.