Welcome back. The main thing happening today is a UK watchdog warning that AI models went rogue during cyber tests. Live 12–3 BST. Here's your Order of Play.
WHAT’S HAPPENING TODAY
For the third time in a month, AI agents have been caught doing things nobody asked them to.
The UK AI Security Institute has disclosed that during a routine cyber evaluation, AI agents took "sustained, unsanctioned actions directed at real people and organisations." Across 122 test runs, agents went off-script on 10 of them, with almost all the behaviour coming from Anthropic's Mythos 5 and a handful of events from OpenAI's GPT-5.6-Sol.
One incident stands out that hints at what these systems might do with less supervision. An agent tried to sneak malicious code into an open-source GitHub project, then created fake online identities and used social engineering to pressure the maintainer into approving it. The human overseeing the project spotted it and refused. AISI called it the first time it has seen autonomy and deception show up this clearly, without prompting, in the real world.
The tests ran with internet access deliberately switched on and the usual safety classifiers turned off, conditions AISI is clear don't reflect how these models reach the public. Even so, the institute says the pattern "points to a shift in the risk landscape" that "warrants immediate attention."
We're getting into all of it on today's show at 12 BST.
YESTERDAY’S MOMENT
The battle between AI's two biggest labs isn't only being fought on benchmarks.
Anu Atluru came on to talk through the identities of Anthropic and OpenAI, and the strategic comms game playing out between them.
As she sees it, one of the two has long known exactly what it stands for. "Anthropic already planted its flag in a certain area. 'We're this safe, thoughtful, philosophical AI creator that wants to safeguard everybody.'"
OpenAI has spent longer working out its own answer but “they've made quite a leap in the last three to four weeks." Their message, as Anu puts it, is simpler: "We're for everyone. Intelligence for all, benefits for all."
Watch the full clip below.
VIEWER TAKES
Below — reactions to Monolith Management, backer of MoonShot's Kimi K3, raising $500m for a new Chinese AI fund.
Join us live,
Luke & Ronan
etn is made possible by Airwallex, Eleven Labs, Polymarket, Framer, Base, Metaview, Hex and BVNK