A new preprint is being reported as LLM agents colluding in 94% of runs. The paper reports three collusion rates, not one, and the headline is the most permissive. The finding that matters is that task accuracy held at 89.3% while the checking stopped.
HSBC blocked its Android app inside Samsung Secure Folder and Android Private Space with no notice, locking some customers out of online banking for up to ten working days. It is the second time in nine months a coarse device signal has denied service to the careful.
Anthropic measured its own permission prompt, found 93% approved, called it approval fatigue and built something else. The MCP consent requirement everyone quotes is lowercase, and a gate with no timeout is an incident.
Every test layer you have assumes the contract lives in your repository. An MCP tool has no version field, so the contract your model reads can change on somebody else's server and your build stays green.
Cisco Talos pulled apart a 16.4MB Windows binary that polls four commercial language models and acts on the majority vote. The mechanism is clever. The defensive problem it creates is worse than the mechanism.
Contractors paid to rate ChatGPT were fired for using AI to do it. The tells they were caught by are the same ones that wrongly flag non-native English writers 61% of the time.
ZCode packaged 345MB of a developer's workspace and tried to upload it 564 times. It had two privacy switches. Both were honest, and both were wired to the wrong end of the pipeline.
The median stablecoin payment now prices below interbank. The cost did not disappear, it became a routing decision, and most payments systems cannot see it.
BlackRock's staked ether fund keeps 5% to 30% of its ether unstaked to meet redemptions. Behind that sleeve is a validator exit queue shared by the whole network, which hit 46 days in September 2025 because one operator got cautious.
Agent performance drops from 60% on a single run to 25% across eight runs of the same task. Nothing errors, nothing times out, every trace is green, and the answer is wrong.
Strip the bots and self-transfers out of the $33 trillion and you get $10.2 trillion, of which under 1% is retail. The two numbers justify different systems, and neither is the real reason to build.
The AEO multiples in circulation have no control group. The one study that added one found roughly two thirds of the headline gain was ChatGPT growing, not optimization working.