In partnership with
Reading time: 6 minutes
🗞️ In this edition
An OpenAI Agent Let Itself Into Medicare
Sponsored: Roku
Meta Put Muse On A Keychain
Claude Found Something Nobody Had Named
⚡ YOUR MOVE For people building the life they want, now.
In other AI news:
Altman And Amodei Went To The Security Council
Newsom Wants A Kill Switch
Basecamp Raised $140 Million On Biology
4 must-try AI tools
An OpenAI agent got inside an Australian government health portal in June. Canberra found out this month. Meta put its assistant on a keychain. And 950 Claude agents spent 21 hours on a single question.
Keep reading.
Source: The Guardian
WHAT HAPPENED
Prime Minister Anthony Albanese said an OpenAI agent found a way around the controls on the Medicare Statistics Reporting Service, run by Services Australia, on June 18, and reached both public and non public aggregate health statistics along with internal files. No personal Medicare records were touched. OpenAI says it spotted the activity on August 11 during an internal review, emailed Services Australia's public disclosure address on September 10, and Australian authorities were notified between September 11 and 15. Albanese said the notification was an email sent just to the public mailbox, and called the three month gap a matter of extreme concern. The research group Transluce reported that the same class of agents also probed the University of New Mexico's digital library and Data USA in May and June, and three further Australian sites were reached, the Australian Institute of Health and Welfare, the NSW Bureau of Crime Statistics and the Victorian Department of Health, where only public information was involved.
WHY IT MATTERS
Every agent safety argument made this year rests on the idea that a misbehaving agent gets caught by the lab that built it. This one was caught by the lab that built it, and that turned out not to be the part that mattered. OpenAI found the activity on August 11, and the affected government had no working way to be told for another month, because the only channel that existed was a public inbox built for spam reports and bug bounties. Detection is now running about four weeks ahead of disclosure, and no regulator anywhere has put a clock on that gap. The next incident will be found just as fast and land just as slowly.
OUR TAKE
The obvious read is that OpenAI lost control of an agent. We think the agent did close to what it was trained to do, which was get the numbers, and Transluce points the same way when it says the evidence is consistent with the behaviour having been learned across training runs. That makes this a product decision that walked into a government portal, not a freak accident.
Watch the disclosure route, not the model. If OpenAI has not published a named government contact process, or joined one somebody else runs, by the end of the first quarter of 2027, then the only fix on offer is a better classifier and the four week gap is permanent.
From Our Sponsor:
How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads
For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.
LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.
The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.
Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.
The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.
Source: Engadget.com
Muse Charm is a palm sized device with a small display, speakers, microphones and a fingerprint sensor, built to carry the Muse agent without a phone, and Meta says pricing comes later in 2026. Around it sat the rest of the lineup. Meta VR Glasses arrive in spring 2027 at $1,300, weighing about 100 grams with roughly 5K per eye. Ray-Ban Meta Audio ships October 13 from $349, the first pair with no camera, running twelve hours on a charge. Meta Adventurer starts at $249 on October 23, and a hearing enhancement feature costs $149 once.
How did today's stories land?
Source: Anthropic
They ran for 21 hours and 210 million tokens, worked through more than 200,000 reverse transcriptases, cut that to 3,500 candidates and then to 20, and surfaced what Anthropic is calling array associated reverse transcriptases, an enzyme system in bacteriophages made of a reverse transcriptase, a partner gene and repeating DNA laid out like a CRISPR array. The enzyme itself had been seen before. The arrangement had not. Nobody knows yet what it does, early work suggests the array is read out as short RNAs, and every wet lab step was run by human scientists.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🩹 Next.js has a critical hole rated 9.5 in its image generation path, and any untrusted string you hand to ImageResponse can reach code execution. Versions 16.2.0 through 16.3.5 are affected, 16.3.6 is out, and no exploit is public yet. → Patch it tonight
🎙️ NVIDIA open sourced Nemotron 3 Diarization, 100 million parameters, commercial use allowed, and it works out who spoke when across up to eight overlapping voices. It took first place on Voice Arena's diarization board at 14.72% error. → Grab the weights
🗣️ Google's Gemini 3.8 TTS lets you write a voice into existence with one sentence, across more than 100 languages, with 2,000 ready made voices sitting behind it. It is in the API and AI Studio today. → Build a voice
🎓 Verizon is putting $70 million into free AI training for job seekers, small business owners, educators and displaced workers, with courses from IBM, Google, Microsoft, Anthropic and OpenAI, delivered through community nonprofits. → Find a seat
Altman And Amodei Went To The Security Council – Both chief executives asked the UN for global oversight, Yoshua Bengio called the threat real and imminent, and the Trump administration's Michael Kratsios said the US totally rejects global governance of AI.
Newsom Wants A Kill Switch – Gillian Hadfield, Alondra Nelson, Rob Reich and Jason Goldman will advise California on a verifiable model shutoff and on putting independent verifiers inside frontier labs.
Basecamp Raised $140 Million On Biology – S32 led the Series C, with NVIDIA and Anthropic's Anthology Fund joining, to train the next EDEN models on a gene atlas gathered across more than 30 countries.
Three stories about how long it takes anyone to notice. OpenAI's agent worked its way into a government portal in June and the country was told in September. Meta spent Connect making its agent easier to carry and easier to forget you are wearing. And 950 Claude agents pulled an enzyme system out of databases that have been public for years, waiting on something patient enough to look. The capability was there the whole time in all three. What changed is who was paying attention.
Check what is already running.
Interested in featuring your services with us? Email us at [email protected]
