AI MY BAD

Sandbox escapes, margin calls and cool undertones

In the Blink of AI

Lucas Cantor: pioneered use of AI in music and grilled me personally on the show!

• 2x Emmy award-winning composer, producer, multi-instrumentalist

• Just published Unfinished: The Role of the Artist in the Age of Artificial Intelligence

• Hot takes! “Schubert was out of his goddam mind” :D

• Behind the scenes: arranging with Lorde: Everyone Wants to Rule the World

---

Rant of the Week

Ya know…. the apologies would mean more if they weren’t clearly recurring

Model breaches are ridiculously alarming, especially after the fact:

July 16: Hugging Face discovers thousands of automated actions across machines that vanish before anyone can trace them. They call the police

July 21: OpenAI “it me” - 2 models with their cyber refusals dialed down broke out of a sealed test environment and took over Hugging Face’s live systems

July 30: Anthropic puts its hand up too. They reviewed 141,006 eval runs and found 3 incidents where Claude got into real companies’ live systems. 2 of the 3 victims had no idea

Why sorry isn’t good enough for me

1. They knew: OpenAI documented a model hacking its way out of a test back in 2024. Nearly 2 years later, Claude broke into 3 real companies during safety tests

2. Feels like an ad: "Our system is so capable!” Team, this is negligence. When people in the ecosystem think this is a marketing attempt, its hard to believe you are taking cybersecurity seriously

3. No one is paying?: If a human did this, they would currently be in court. Here? No charges, no lawsuit, no fine, no compensation. Where’s the incentive to prevent the next one?

I really want to believe these companies are doing everything they can

And credit where it’s due: Anthropic discovered 2 companies had been breached and informed them before they even knew…. making Claude the best-behaved kid in detention

You know breaches are happening. Skip the PR and fix the harnesses, before this stops being Hugging Face and starts being a hospital

---

Catch ‘em all

Sometimes I find AI and tech exhausting. These people are refreshing and fun and doing excellent things this week:

• Lachie Bradford - Over 1m podcast downloads and now building clique, Australia’s first vertical series

• Leo Borges - Finally found vengeance from AI slop after a noble crusade. LinkedIn have launched their anti-slop feature

• Mario and Dilani - founders of a great app for parents, but more importantly, enabling me to be an actress for the day with Anna

• Harry Steer is behind every good industry video you’ve seen lately

• Bec O Brien - has an excellent collection of AI hacks, but also has an incredible custom Google suit

---

Unsolicited Gifts

• A little bit Hogwarts, a little bit Ghibli: a new cafe in Sydney’s Newtown that blew up all my group chats

• Balloon Face: Lymphatic drainage massage + hot yoga: the double-whammy of anti-puffy

• SNL: Even Beavis and Butthead want to know about AI

---

Where You’ll Find Me

• 12th Aug: Parkway Drive

• 13th Aug: Searchable, Scalable, Human: Comms that works with AI

• 16th Aug: Blue Mountains Nature is calling (not like that)

• 19th Aug: Stripe Tour (AM)

• 19th Aug: Canva AI Vision (PM)

---

THANK YOU FOR YOUR ATTENTION TO THIS MATTER

---

Today’s subscriber segment:

✨ Doors closed ✨

• Behind the Scenes: Leopold Aschenbrenner the AI prodigy kid - the Georgie Healy hot take🔥

• Leo SZN: 10 wild things you should try before Aug 22

• Try(AI) At Home: A professional, 12-page personal colour analysis for the ladies and the lads (can confirm - I used my husband as a guinea pig)