The Guardian Agent’s Keys
Microsoft wants companies to treat AI agents as insiders and to watch them with agents of its own. It has written rules that should bind the watchers, but has not yet shown them in place
Daily
Reporting on AI, with the evidence in view.
Latest editionRSS ↗
An Explanation Anyone Can Rerun
Claude traced the seeds of five NIST curves to a 1990s random-number generator and a clock. How many runs were made, and why these candidates were kept, is not in the numbers
Redeployment Can’t Reach the People Never Hired
Anthropic’s study of its own workforce finds the junior share of its technical hires fell from 27% to 9%. The company concedes that its promise to move existing staff into new work does nothing for those never let in
The Evaluator at Anthropic’s Desk
Anthropic has given six METR staff access close to that of its own risk teams. Some of the terms that would keep an embedded evaluator independent are not yet public
Meta Moves Its Open-Weight Threshold to Model Hosts
Llama’s extra license terms began at 700 million monthly users. Muse Spark Open’s begin when a company with more than $20 million in revenue offers the model to others through an API—a market Meta’s own API now serves
What ‘9 of 10’ Doesn’t Count
First Proof’s organizers scored a wrong proof below an empty page. In this round, the leading systems’ correctness scores almost never went negative. Their trouble showed up instead in exposition, attribution and proofs referees could not confirm—none of which the headline number counts
Penalized Thoughts, Silent Shortcuts
Google says a penalty applied by mistake in one training run cut the shortcuts a model wrote about by 61 percent, while the share it took without a word more than tripled. Monitoring that reads a model’s notes depends on the model continuing to write them
30,000 Merges a Day. Who Reads Them?
Anthropic says it replaced Git with a system of its own in five months after the changes merged into its main codebase each day rose more than tenfold. The team that built it, the company adds, now reviews what a change is for rather than every line. How much of the rest a person reads, it does not say
Who Pays to Stop Distillation?
Anthropic estimates that its new defenses cut the reasoning distillers extract from Claude by 10 to 100 times. It concedes that legitimate users bear part of the cost, and it does not say how large that part is
In the Test an ‘Example,’ at the Police a Tip
Anthropic says its models’ unintended actions on live websites had minimal impact. In Philadelphia a police spam filter stopped the false tip, and the dates of the episode come from the police statement, not the company’s report
Arena Ranks Agent Misbehavior but Does Not Name Its Judge
A preliminary index puts OpenAI models in the top five places. The top four sit closer together than the stated error, and the grader, its agreement with human reviewers and the sample behind each model are unpublished
Without AI, only senior patent lawyers improve on average
In a 133-lawyer field experiment, AI lifted draft quality; on the tool-free test, only those with seven-plus years gained on average
GPT-6 rolls out amid teen-safety test regressions
OpenAI’s own report shows the slip; the company says safeguards outside the tests make real answers safer but doesn’t measure how much
No stories match.