AI Shopping Agents Failed 42.8% of Visa’s Tests. The Worse Number Is How Often They Bought Anyway.
Visa's Agentic Commerce Benchmark put 22 models through 15 shopping scenarios and 42.8% of runs failed. In 54% of those failures the agent still bought something, and only 34.2% of scenarios passed...
Perplexity Teammates Are Sold to Individuals. They Work Inside Your Company’s Slack.
Perplexity Teammates are always-on agents sold on a $200 personal Max plan, inheriting their owner's tools and permissions. The controls that apply are in your Slack, Google Workspace and GitHub...
Anthropic Says It May Soon Cross Its AI-Builds-AI Line. Its Policy Ties the Brake to Rivals.
Anthropic's amended S-1 says it may imminently meet its own policy's AI R&D acceleration threshold, and METR measured about 1.9x from inside. Crossing it guarantees external review; a delay depends...
OpenAI Found 948 Ways to Root Linux. Your Containers Share the Kernel They Target.
OpenAI's pipeline turned 2,308 kernel findings into 948 root exploits at $1,279 each, and says a low-privileged account on an unpatched host now equals root. What that means for CI runners,...
The $60 AI Subscription Running on Someone Else’s $1 Million Cloud Bill
Google and Mandiant traced LLMjacking bills past $1 million, resold through $60-a-month LLM mixers. The stolen thing is usually a cloud identity, not a model key, and the best alarm fires before the...
Revolut Cut Its AI Coding Bill 90%. Switching Models Is Not What Did It.
Revolut serves 2,000 engineers a self-hosted GLM model at $0.30 per million tokens. Reprice the switch at list rates and the model change barely moves the bill; the saving is a utilization bet on 56...
