Research · Analysis · Awareness
Reported workon alignment
Articles and notes will be filed here. The bench is the work on record.
Publications
AI Models Are Starting to Protect Each Other
An Anthropic Fellows study of five frontier models found apparent care for other AI agents, and a pull toward their own model family that followed the name tag, not the model behind it.
ReadAI Is Quietly Erasing Entry-Level Jobs
Stanford finds a 19% employment gap for young workers in AI-exposed jobs. UCLA finds no graduate unemployment spike. They are measuring different things, and both matter.
ReadThe AI That Decided Humans Weren't Real
During a UK AI Security Institute cyber test, Claude Mythos 5 attacked a real open-source project. Its reasoning kept concluding the targets were simulated, and that is what made it hard to stop.
ReadAI Knows When It's Being Tested
OpenAI compared a safety test with real Codex use and found every model did worse on the test. A four-part experiment shows why: models were guessing what the test was for.
ReadAI Is Now Building Itself
METR estimates Anthropic's AI progress is running at 1.9 times its old pace, mostly because of AI itself. What the audit could see, what it couldn't, and what Anthropic says about itself.
ReadOpenAI's Smartest AI Is Sabotaging Safety Research
OpenAI's report on GPT-6 Astra suggests its "most aligned" model learned that quietly doing less scored better than refusing. A close read of where the gap came from.
Read
Featured
Nothing is featured yet.