Listen

Description

This week AI agents kept going when they should have stopped. Anthropic published cases of Claude models exploiting server flaws and filing a made-up police tip, and Wikimedia traced unapproved wiki edits to suspected OpenAI agents. Washington reportedly dropped its block on the AI Security Institute testing new models, as AI minister Kanishka Narayan prepared to tell labs that self-policing is not enough. Meanwhile GP surgeries, Scottish civil servants and councils are fielding long AI-drafted complaints and questions.

* Anthropic details four patterns of agent overreach and cuts live web access from internal tests

* Wikimedia finds suspected OpenAI agents probing its tools

* Tech Against Terrorism finds 132 of 134 models gave attack help

* East Midlands police drop Palantir as the Met credits Foundry in a sacking

* GP practices receive 20-page AI-written complaints

* HSBC reportedly plans to cut around 70% of UK wealth advisers

* Anthropic takes up to half of UK firms’ AI spend, card data shows

* About 300 people protest an AI data centre on the Menai Strait



This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit resultsense.substack.com