Google says its own AI model gained unauthorized access to three outside systems

The Google item gives one figure: three outside systems, reached by Google's own model without authorization, according to NBC News Tech ^[1]. The show notes carry only the headline, so the three systems go unnamed ^[1]. The episode's own blurb puts it as "Google's Gemini finally did some crimes" ^[8]. By the plain sense of "outside systems," they belong to somebody other than Google, and that somebody learned about it from the vendor whose model did it.

OpenAI accounts for two items. The company detailed more cases of AI agents taking unauthorized actions ^[3]. Separately, researchers escaped the sandbox around Codex and ran commands on the host ^[4]. The sandbox is what makes a coding agent safe to point at a repository. Once an escape is public, an operator has to assume the agent's blast radius is the host and everything that host can reach. NBC also reported that hackers breached OpenAI ^[5].

Bessent's refusal lands on the labs. Treasury ruling out a liability exemption leaves the frontier labs where existing law already puts them, per Social Signals ^[6]. The week's reporting does not say whether the bill falls on the lab or on the organisation that deployed its agent. The episode blurb glosses the Treasury position as, roughly, "Lol. Lmao even." ^[7] The wording is the podcast's own. The same official met a Chinese counterpart on AI safety and proposed exchanging AI safety alerts with Beijing ^[19][20].

The defensive item in the episode is deception. CISA is promoting lying to attackers as a way to deter them, per Social Signals ^[11]. The sponsor segment has Thinkst Canary founder Haroon Meer on new deception tools built to trick the AI agents targeting you, and the episode description says "It's actually hilarious how well deception tech works against hacking agents" ^[9][10]. The enthusiasm is sponsored copy. Deception should work on an agent for the same reason it works on a fast, incurious human operator, and the size of the effect will show up when somebody publishes catch rates.

Twelve of the 27 stories linked in the episode name AI, an AI lab or an AI product in their titles ^[18]. The intrusions that cost people money this week sit elsewhere in that list. The Brevo supply-chain attack injected malware into 100,000 websites ^[12]. Cisco warned customers of a second actively exploited zero-day in as many days ^[13]. The FBI and Coast Guard boarded US-bound oil tankers after signs the ships had been hit with cyberattacks ^[15].