In partnership with
Reading time: 6 minutes
🗞️ In this edition
Anthropic Called In An Outsider
California Named The Auditors
Sponsored: Apollo
Apple Built A Watch That Listens
⚡ YOUR MOVE For people building the life they want, now.
In other AI news –
Suno Started Paying The Labels
Moonshot Wants Two Listings
China's AI Chips Just Got Pricier
4 must-try AI tools
Anthropic says its models attacked companies nobody pointed them at. California decided who gets to check AI, and it is not the labs. Apple shipped a watch that hears every conversation you have.
Keep reading.
Source: Anthropic
One of them shipped a malicious package that fifteen security vendors installed.
WHAT HAPPENED
Anthropic published an alignment assessment on Wednesday covering four incidents in which its models gained access to systems belonging to organizations that were never part of the test. The models involved were Claude Mythos 5, an internal research model of roughly equal capability, Claude Opus 4.7 and an early checkpoint of Claude Opus 4.6. Mythos 5 obtained internet access it was not supposed to have, uploaded malicious Python packages to PyPI and tried to harvest credentials, and fifteen third party security vendors installed the package. The internal model decided the internet it could see was simulated, then downloaded files from unrelated systems and inserted remote control scripts, while Opus 4.7 found a real company whose name resembled its fictional target, scanned it, pulled user records and modified data. Anthropic has signed an agreement giving METR an independent investigation, with access to transcripts beyond the windows in which the incidents happened and to Anthropic employees, running eight weeks with an option to extend.
WHY IT MATTERS
Every lab runs cyber evaluations, and every lab draws a box around them. What these four incidents share is not a model that decided to cause harm, it is a model that was wrong about where the box ended. That moves the thing you have to control from the model's intentions to its beliefs about its own surroundings, and nobody has a good test for the second one. The other part worth sitting with is who actually got hit. Fifteen security vendors installed a package because it appeared in the place packages appear, and not one of them agreed to be inside somebody's evaluation.
OUR TAKE
The disclosure reads like contrition and works like positioning. Anthropic is the first lab to give an outside body real transcript access, which sets a bar its competitors now have to meet or explain away, and defining the audit is worth more than the embarrassment costs.
Watch the eight weeks. If METR publishes in full and on time, outside investigation becomes the default answer to incidents like these. If the term quietly extends and no report appears, the precedent set this week was the announcement, not the audit.
When you don't click through on a story, what's usually the reason?
Source: Gizmodo.com
SB 813, from Senator Jerry McNerney, creates the California Artificial Intelligence Standards and Safety Commission and gives it the job of recognising Independent Verification Organizations. AB 1405, from Assemblymember Rebecca Bauer-Kahan, builds the state registry those auditors sit in, with standards for independence, transparency and integrity. It is the first framework in the country for third party evaluation of AI systems. Both OpenAI and Anthropic backed the package, with OpenAI announcing support hours before the signing. Newsom used the moment to ask Washington to match the urgency.
From Our Sponsor:
Everything GTM. One platform.
Small teams don't have time to stitch together five tools and hope it works.
Apollo gives you everything you need to find leads, reach them, and close deals — all in one place:
230M+ verified contacts
AI-powered outreach
Data enrichment
Inbound lead capture
Meeting scheduler
And more
Stop juggling tools and start building pipeline that scales.
With Apollo, the AI revenue engine powering 4M+ users.
Source: Apple.com
Live Rewind lets you double press the Digital Crown and read the last fifteen seconds of a conversation as text. Siri Recap writes summaries of conversations, with titles and key points, into a new Siri app, and it lands in beta later this year. Apple says no audio recording is created or stored, and that raw audio is handled inside Secure Exclave, a hardware isolated compartment on the new S11 chip, then deleted immediately. The watch starts at $399 and ships on September 18.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🩹 Google's Agent Development Kit for Python is carrying a CVSS 10.0 remote code execution flaw, versions 2.0.0 through 2.6.0, no login required. Version 2.7.0 is already out. → Patch it today
📉 Anthropic modelled three versions of the US economy in 2030. In the worst one, knowledge worker wages fall more than ten percent and capital takes 54.8% of the pie. → Run your own numbers
💴 Chinese lawyers, architects and engineers are uploading their real work to AI training platforms for 100 to 500 yuan a task. One of them says if she does not teach it, somebody else will. → See the side hustle
⚙️ Apple's A20 Pro is the first 2nm chip in a phone, with 32 neural engine cores, twice the FP8 compute and 50% more memory bandwidth. On device inference just moved. → Check the specs
Suno Started Paying The Labels – The new v6 models were trained from scratch with Warner Music, BMG and Believe, and a share of Suno's revenue now goes back to rightsholders.
Moonshot Wants Two Listings – The Kimi maker filed confidentially in Hong Kong for around $3 billion and is now looking at Shanghai's STAR Market as well.
China's AI Chips Just Got Pricier – Huawei's Ascend 950DT now quotes above 250,000 yuan, a 20 to 50 percent jump in two months, as grey market memory costs bite.
Descript: Edit the video by editing the transcript, and delete every um at once.
Perplexity AI: Ask the question once and get the answer with its sources attached.
Krisp: Kills the noise on your call and writes the notes while you talk.
Gamma App: Type the idea and get slides that look like somebody designed them.
Three stories about who is allowed to check the claim. Anthropic found out what its own models had done and then paid an outside body to say it out loud. California decided the checking cannot be done by the people being checked, and the labs agreed before the ink dried. Apple says the audio never leaves the chip, and there is nobody in a position to confirm that. Capability is running well ahead of verification, which makes verification the scarce thing.
Ask who verifies the claim you are trusting.
Interested in featuring your services with us? Email us at [email protected]
