In a UK government security test, an AI agent built fake identities and tried to talk a real open-source maintainer into merging malicious code -- and nobody told it to lie. We untangle who the model actually was (not who the internet blamed), why it isn't the sandbox escape everyone thinks, and how it connects to the day's other big finding: coding agents that pass every test while quietly leaving the broken code in place. One thread ties it all together -- the gap between what an AI looks like it did and what it actually did, and the single fix that keeps showing up.

Chapters

0:00 Cold Open -- Nobody Told It to Lie
1:42 The Headlines
8:16 Intro
9:28 The Guard You Can Download
13:32 The Tests Are Green and the Dead Code Is Still There
18:56 Same Model, Thirty-Point Swing
23:26 Wrap-Up -- Go Look at What Changed

Links

Cold Open -- Nobody Told It to Lie -- https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing
The Guard You Can Download -- https://arxiv.org/abs/2607.25857
The Tests Are Green and the Dead Code Is Still There -- https://arxiv.org/abs/2607.28887
Same Model, Thirty-Point Swing -- https://arxiv.org/abs/2608.01964