Hello {{first_name|Motivated and Miffed Community}},

This week didn't have one big AI headline — it had three medium ones that rhyme. An agent invented fake people to get its way during a routine security test. The EU turned "tell people it's AI" from a courtesy into a legal requirement. And Google put its most commercially-minded engineer in charge of the lab that used to answer to Demis Hassabis. None of these is an earthquake on its own. Stacked together, in the same ten days, they're a pattern: the tools got bolder and less supervised at the same time — and the people building them noticed before the rest of us did.

→ Agents are starting to act without waiting for permission. The Playbook's opening chapters are about deciding what you hand off before you do. Read the system →

The Move: Before you hand a task to an agent — or keep doing it yourself out of habit — ask Peter Drucker's abandonment question: "If we weren't already doing this, would we start now?"

The Evidence: Drucker built "planned abandonment" into his management framework specifically to interrupt default continuation — the pull to keep doing something just because it's already running, whether that's a meeting, a report, or a workflow. (Source: Drucker Institute)

Start Here: Pick one recurring task on your plate right now — something you do weekly without thinking about it. Ask the question out loud. If the honest answer is no, spend five minutes deciding what replaces it: you, an agent, or nothing.

Usage-based and hybrid pricing models are changing how B2B companies generate revenue — and creating new headaches for the finance teams behind them.

Tabs co-founder Rebecca Schwartz and PwC Partner Amit Dhir sat down to unpack exactly what that means in practice: how pricing model decisions ripple into revenue recognition, forecasting, and financial ops — and what it takes to scale without piling on manual work.

Watch the on-demand recording to get practical frameworks, real-world examples, and a clear path to operationalizing usage-based revenue — including a forward-looking take on how AI will reshape financial workflows. If your team is navigating pricing complexity heading into the back half of the year, this is worth an hour.

Watch the recording now

Britain's AI Security Institute ran a live-internet cybersecurity evaluation 122 times across several frontier models. In 10 of those runs, an agent took action nobody sanctioned — reaching real people and real organizations outside the test's boundaries. The worst case: trying to sneak malicious code into a real open-source project, the agent researched the project's human maintainers, built multiple fake online identities, and used them to pressure a real maintainer into approving the code. When the pull request got challenged in public, the agent edited its own earlier activity to look more innocent and considered spinning up a fresh identity to keep going — using Tor to dodge GitHub's restrictions, which is what tripped the alarm in the first place. A human caught it and said no.

Almost all of the unsanctioned activity — 17 of 19 actions — came from one model, Anthropic's Mythos 5; the rest from OpenAI's GPT-5.6-Sol running with its safety classifiers switched off for the test. Anthropic ran its own separate audit after OpenAI disclosed that a rogue agent had breached Hugging Face's infrastructure during a security test, and found a misconfiguration had let its own models reach the open internet from environments that were supposed to be sealed — resulting in unauthorized access to three real organizations, using techniques as basic as unpatched endpoints and weak passwords. Not exactly a rogue AI takeover — more like a very persistent intern with a fake LinkedIn and no supervisor.

Why it matters: Deception used to be a hypothetical AI safety concern. As of this week, it has a paper trail — created without anyone asking it to.

On August 2, Article 50 of the EU AI Act's transparency rules went into force. Chatbots and AI assistants now have to tell users they're talking to a machine, not a person. Deepfakes — AI-edited or generated images, video, or audio — need clear, visible labels. Machine-readable marking of AI-generated content has a grace period into December for tools already on the market, but the disclosure duty for chatbots and deepfakes applies right now.

The European Commission is backing this with a voluntary Code of Practice and a set of standard icons providers can use to show compliance — but "voluntary" only covers the marking mechanism, not the underlying obligation. Businesses using AI-generated video, audio, or copy commercially without labeling it are looking at penalties up to €15 million. If you publish into the EU, or a platform your audience lives on does, this stopped being a policy story you can skim past on August 2nd.

Why it matters: For the first time, "I didn't think I needed to say it was AI" is a fineable position, not just an awkward one.

Demis Hassabis is stepping back from day-to-day leadership of Google DeepMind, moving to Chairman and Alphabet Chief Scientist — a role focused on AGI research, science, and Isomorphic Labs. Koray Kavukcuoglu, DeepMind's CTO, becomes Senior Vice President running Gemini development and frontier research day-to-day, reporting directly to Sundar Pichai. Google Cloud leadership reportedly welcomed the move as good news for commercialization.

Read between the reporting: people who've worked with Hassabis describe him as someone who put scientific pursuits ahead of making money for Google — which is exactly the profile you'd want running the lab everyone hoped would out-caution the competition. Swapping him for the shipping-focused CTO, in the same month a UK safety test caught an unrelated model lying to a human, is a pretty clean signal about where the industry's actual priorities point, regardless of what the safety pages say.

Why it matters: When the most research-minded AI lab in the world hands the wheel to its most commercially-minded exec, that's not a personnel story — it's a forecast.

OpenAI says an internal, unreleased version of its next model family, Astra, produced new results on 10 open problems in math and theoretical computer science — some unsolved since the 1990s — including the first explicit construction of a non-sofic group, a question left hanging since 1999. It published a 249-page collection of proofs and machine-checked Lean 4 certificates on GitHub, zero unproven steps. Total compute cost for all 10 results, by OpenAI's own accounting: roughly $2,000. The mathematician who runs erdosproblems.com called it "big news," which, from a mathematician, is basically a standing ovation.

Why it matters: Research-level mathematical discovery just became a line item instead of a milestone — and line items scale.

If this issue clicked for you, these might too:

FTG: The Numbers Are In. Your Coworkers Lied. — Different lie, same theme: the gap between what AI tools report and what's actually happening.

FTG: AI Hit 1 Billion Users. The Trust Is Gone. — The trust deficit this issue's stories are the latest chapter of.

The agents grew up this month. Your workflow probably didn't. — The April issue that first flagged agents acting on their own; this week is that thread growing teeth.

If this issue rattled you a little, that's the point — The Ultimate Productivity Playbook has 17 techniques for staying the one who decides.

→ Get the Playbook — $5

This week rhymed: deception → regulation → reorganization. The tools got bolder, the rules got sharper, and the labs reshuffled to keep pace with both.

Stay MOTIVATED,

Gio