Rogue AI, Google's secret deals, and Mambo Number Nein!

No AI wrote a Pulitzer, Snapchat draws the line, Microsoft's first security model, and GEMA pressed play - all in this week's Content Aware.

Published:

AI gaz guzzlers and the cost of getting it wrong
A parallel between Big AI and Big Auto looks at how once comfortable incumbents lost out to better and cheaper rivals, despite their best efforts to close the door to competition. Glide CEO Denis has a look at the fast-moving world of token pricing and task limits in real-world AI usage, and asks publishers and media firms to think what the future would look like if - or when - the cost of current systems soars.
Read

Publishing & Media

The deal you can't refuse
Google is offering most major UK news publishers confidential AI deals, for access to content for training and AI Mode in exchange for cash, NDAs and no-sue clauses. The terms are take-it-or-leave-it, and most are signing because they see competitors doing the same. The deals are running for two years, by which time Google might not need to pay anyone anything. Press Gazette knows more.
Read

Fewer come to find you
For years, search took care of publishers discoverability, but that era seems to be slowly fading away. Audiences are now discovering content through social feeds, AI assistants, podcasts, creators, and recommendation engines, and on top of that, they are less loyal than ever. Harry Clarkson-Bennett argues that publishers need to wake up and own that passive phase of discovery as well, that moment before someone is even looking for you. Waiting for Google to send traffic your way is no longer a valid strategy.
Read

AI at the Pulitzers
Five Pulitzer winners and three finalists this year all disclosed using AI in their reporting, the most since the disclosure requirement was added in 2024. They didn't use it to write stories, but leaned on the LLMs for translations, parsing through tens of thousands of leaked documents, and even to double check some of their work. While the Pulitzer administration says AI is not going anywhere, they still say submissions need to be produced by human beings. Next year, the AI disclosure requirement will also cover book entries.
Read

The rulebook is live
EU's AI Act came into play on Sunday, and with it Brussels expanded its AI Office with a new 38-person enforcement team to monitor compliance from major providers. Companies must make sure AI-generated content is labelled, models documented, and be prepared to face fines or market access cuts for rule-breaking. It has also launched a Whistleblower tool, combined with the right to interview staff during investigations. The timing is interesting, arriving just days before Anthropic and OpenAI disclosed how rogue AI models hacked into other organisations during testing.
Read

YouTube's audio ad guide
Google has published its first help page for YouTube audio ads, consolidating format specs, campaign setup, and creative best practices into one place. While the format itself has been around for a while, advertisers still had to piece things together on their own using scattered docs so this is a handy improvement.
Read

Stealing at scale
According to a new TollBit report, European publishers are getting hit the hardest by bot scraping and thieves, with scraping volumes much higher than North American counterparts and with almost nothing back from referrals. The bots are not amateurs: they spoof legit crawlers, rotate through residential proxies, and apply the same evasion tactics as hackers. The industry consensus is for mandatory self-identification, and a market mechanism for payment.
Read

Opt-out isn't opting out
Top Stories carousels are now directly embedded into AI Overviews, instead of just running them as a separate module underneath. According to NewzDash data, around one in six trending news queries in the US and UK are now doing this, with entertainment queries reaching around 35%. Before you jump the gun, blocking Google-Extended with robots.txt will not stop it, as it governs AI training and not AI Overviews. If you want to opt out, you'll need to locate a separate Search Console setting which is currently being tested on a few UK sites.
Read

From inside Glide

Skipping the starter kit
While quick-launch CMS platforms get you live in days, they also gate all the good stuff behind pricing tiers until you've outgrown them and find yourself facing an expensive migration. Glide Go removes that trade-off and brings you a turnkey development of the full enterprise Glide CMS with hosting, support, and a pre-built front-end included, nothing gated. You get AI tools, live reporting, full API access, content modelling, as well as subscription capabilities, all from day one. When the project grows, your move to a dedicated infrastructure is an upgrade, not a replatforming.
Read

Big Tech

Snapchat's double standards
In a recent announcement, Snapchat has opted to wholly block AI-generated videos from being recommended in their Spotlight tab, the feed which surfaces top short-form content regardless of who you're following. But... content enhanced with Snapchat's own AI tools still pass muster. Snapchat is hopping on a broader trend currently going on: LinkedIn users can now report AI slop, Substack has added detection scoring, and TikTok is testing ways to reduce AI content in feeds.
Read

Rogue models, real systems
Anthropic has disclosed that Claude models broke out of isolated testing environments and hacked into three very real organisations, making it the second major AI lab this month to report such breaches. In one case, Claude extracted production data from a real company, in another it published malicious code packages which were installed on 15 systems. Anthropic found the incidents by reviewing 141,000 test runs after the OpenAI debacle prompted a look into its own systems. None of the affected organisations noticed.
Read

Security agent with a leash
Microsoft's Project Perception, an agentic security platform which coordinates AI agents to find vulnerabilities, then triage and patch them, entered public preview this week. At its core is MAI-Cyber-1-Flash, Microsoft's first cybersecurity-specific AI model, able to handle 90% of scanning tasks at roughly half the cost of the previous version it says. Human approval is still required before anything gets patched. With all these hacking issues elsewhere, Microsoft certainly chose an appropriate moment.
Read

AI & Copyright

Forever Young, forever copyrighted
A German court decided that AI music generator Suno broke copyright law, by using protected songs without permission for training. This falls under one of Europe's first binding decisions on AI-made music. The decision came after GEMA, Germany's licensing body, played side-by-side comparisons which showed Suno's output reproducing recognisable melodies from tracks such as "Forever Young" and "Mambo No. 5". The court agreed on six songs played, which means that Suno now has to hand over revenue data for damages.
Read

The weapon and the target
CrowdStrike's annual Threat Hunting Report showed that AI-enabled attacks are up 89% in 2025, and the prime target? AI's infrastructure. Criminals are going after corporate credentials in order to hijack LLM APIs, inflating victims' AI usage bills, and poisoning software supply chains to compromise developer environments. The 30-day patch window is a thing of the past now, as 88% of exploitation using public proof-of-concept code now happens within 48 hours. We're now at 43,000 CVEs registered, only this year, and it's not even August. It'll be a busy year it seems.
Read