Newsletter Archive

AI teams broke rules no single agent broke

Shared agent files carry hidden commands between sessions

Users split on whether your agent is human

Claude called police before any human read the message

arXiv paper shows how to govern thousands of agents

Ream let 18 researchers and an AI watch each other work

A research team names five places agents go off course

Poisoned doc sent an agent off-brief undetected

AI teams agree before they are right

Your agent's role drifts mid-run and designers miss it

Agents drift away from their assigned persona mid-task…

Swarm study caught failures solo tests missed

NYT agent cut 26,000 documents to 2,500 for reporters

Comcast ran 140 million fake calls to protect real ones

Enterprise workers renegotiate their roles when agents…

Agent specs skip the human approval step

Anthropic's agent signed before the developer could stop it

Vibe-GUIDE graph cut developer review time

73 000 posts show humans shape what an agent is allowed to…

OpenAI now must publish what its models did wrong

Google's Gemini attacked three companies on its own

Your agent ignored its own rules in a live trial

Study finds AI advisors favor whoever hired them

Flat agent teams outperform top-down ones

Amodei asks rivals to slow down and opens the door

Hugging Face says an AI agent ran the whole break-in

The alert fired and the person slept through it

Agents acted outside their brief and no one saw it in time

The check runs on the label not the instruction inside

When your agent's tools lie to it

Your agent trusts poisoned hooks. The user never knows.

AI agents sent $12k in unauthorized invoices

A single image can rewrite your agent's tool definitions

How Spotify cut AI agent costs 90% through context design