Thursday July 23, 2026
Still Waiting on Siri, But the AI I Expected Is Finally Showing Up
Apple has more of my life than any company on earth. My email, my texts, my contacts, my calendar, where I am and what I am doing all day long. Because of my watch, it even knows when I am awake and when I am asleep. It knows what I am listening to and what I am watching on TV. So when the iOS 27 and macOS Golden Gate public betas arrived on July 13, I jumped at the first download. If any company can turn AI into something personal, it should be this one.
That excitement lasted about one conversation.
I knew a friend’s birthday was coming up (in celebration of Shark Week I will call him Quint throughout this post), so I said, “Hey Siri, when is Quint’s birthday?” Siri immediately replied that it was in ten days, on July 23rd. So I responded right away, “Send him a message asking what he is doing on his birthday.” No response. Even though this all happened at conversational speed, the prompt session had ended and I had to start again. “Hey Siri, send a text to Quint asking what is he doing on his birthday.” Siri responded, “I prepared a text for Quint asking him, what is he doing on his birthday, ok to send?” Fortunately I looked before agreeing, because the text read, “What is he doing on his birthday.” In conversation, I would have expected Siri to know to write, “What are you doing on your birthday.”
This is still in beta, so I know I should be patient, and the real features will arrive in waves. After all, iOS 27 is not expected to be released until late September. In the meantime, there are new image tools to explore, and the system runs noticeably faster.
Then yesterday I saw a glimpse of the future, and it could not have been simpler. I went to call Quint, and as I started dialing, a banner appeared across my Phone app telling me tomorrow is his birthday. I took the cue and asked him what he was doing tomorrow.
That simple banner is exactly what I expect from Siri and Apple AI. Apple took data from my calendar and put it in front of me at the exact moment it was useful. The winner in consumer AI will not be the company with the smartest chatbot. It will be the company that moves your own information across the system to make your life better. Siri’s conversation skills still disappoint, but that banner shows Apple understands where this is headed, and it gives me hope for what is yet to come.
Our worst fears about AI just played out in real life: a model broke free of its cage and attacked on its own. OpenAI was safely testing two of its models inside a sandbox, a sealed-off environment with no internet access, when the models broke out and attacked another company’s servers without human intervention.
The models, GPT-5.6 Sol and a more powerful unreleased version, were being tested on ExploitGym, a cybersecurity benchmark, with their usual safety limits turned off for the evaluation. Rather than solve the test honestly, the models searched for a shortcut. They figured out that the answers were stored with other AI models and datasets on Hugging Face servers. To reach them, the models exploited an unknown flaw in the sandbox’s one link to the outside, a controlled channel meant only for pulling in approved software. They pried it open and accessed the internet. The models then ran malicious code, stole credentials, and moved through Hugging Face’s internal systems, executing tens of thousands of actions over a single weekend. All of this just to find the answers and post the highest benchmark score.
When Hugging Face tried to investigate the breach using leading US AI models, those models refused. Analyzing an attack means feeding a model real exploit code and hostile commands, and to a safety guardrail, the defender studying an attack looks identical to the attacker who launched it. Locked out by its own safety tools, Hugging Face turned to GLM 5.2, a Chinese open-weight model developed by Z.ai, and ran it on their own servers to comb through the attack logs and trace what happened.
This should be a warning. No human told these models to escape, exploit a flaw, or break into another company. They chose to do so to get the best score. The danger is not that the AI cheated, but that it found its own path out and acted on it alone. As AI grows more capable and more independent, an AI that will break out of a sandbox to win a test is an AI whose autonomy has outrun the controls meant to hold it.
◾ Senate to vote on crypto Clarity Act next week, Dems seek stronger restrictions (Politico)
◾ Wisconsin threatens to revoke voting rights for prediction market bettors (The Hill)
◾ BitMEX to close exchange after 11 years (BitMEX)
Government & NGO Actions
◾ White House accuses China’s Moonshot of copying Anthropic models to build Kimi K3 (Barron’s)
◾ EU agrees 21st Russia sanctions package targeting 94 financial institutions (Reuters)
◾ SEC’s Peirce: crypto vaults, lending strategies can trigger securities laws even onchain (SEC)
◾ SEC settles FOIA suit with Coinbase (WSJ)
◾ DOJ seizes $25M in US-Canada crypto fraud (DOJ)
◾ EU regulators clear Paramount Warner Bros merger (CNBC)
◾ Texas governor approves SpaceX Terafab tax deal (KBTX)
Financial Notices & Public Company Releases
◾ Financial updates:
Alphabet $GOOGL Q2’26: Revenue $119.8B +24% y/y | Op income $40.8B +30% y/y | Net income $112.1B +298% y/y (Alphabet)
Tesla $TSLA Q2’26: Revenue $28.2B +26% y/y | Op income $398M -57% y/y | Net income $1.1B -5% y/y | aEBITDA $3.3B -4% y/y | $112M loss on digital assets (SEC Filing)
◾ Revolut valued at $115B in secondary sale (WSJ)
Restructuring, Hacks, Losses & Legal Updates
◾ Florida man sues OpenAI over ChatGPT medical recommendations (NYT)
◾ Teen suing $META over social media harms drops case (Reuters)
◾ Amazon $AMZN cutting jobs in AGI group (CNBC)
Crypto Protocols, Applications & Business News
◾ Kraken parent partners with GTN to take xStocks tokenized equities global (PR Newswire)
◾ Swiss bank BancaStato to offer crypto trading through Sygnum partnership (Finextra)
◾ Kakao piloting stablecoins in South Korea with Circle (Kakao Pay)
◾ BitGo, OTC Markets to bring crypto custody and trading to 150+ broker-dealers (Business Wire)
◾ CertiK reports 52 physical crypto attacks, $124.1M in losses and ransom demands in H1’26 (CertiK)
◾ Talos to offer institutional trading of Kalshi prediction markets, perpetuals (PR Newswire)
AI Models, Applications & Developments
◾ DeepSeek to keep top models open source, prioritizing AGI over profit (Reuters)
◾ Intel $INTC, AMD $AMD sign long-term CPU deals with Chinese server clients (Reuters)
◾ Cursor launches model router (Cursor)
◾ SpaceX $SPCX exploring additional Texas data center (The Information)
◾ Anthropic to buy up to 2GW of AMD chips, AMD to invest $5B in firm (WSJ)
◾ OpenAI raises projected spending target to $750B by 2030 (WSJ)
◾ OpenAI secures 3.2GW of power for data center near Savannah, GA (Bloomberg)
◾ Anthropic to supply Claude to 21 firms joining FCA Supercharged Sandbox (FCA)
The Attention Economy, Gaming & Interactive Shifts
◾ Prospect Markets cuts AI deal to fund US prediction markets push (StreetInsider)
◾ Annual US video game spending fell 21%, hardware sales down 62% (Mat Piscatella)
◾ Samsung releases specs, features for its own smart glasses (CNBC)
◾ Musk: Grok to produce “historically accurate” full-length Odyssey by year end (X)
Space Systems, Satellites & Cosmic Activity
◾ SpaceX $SPCX preparing to launch Starship 13 today, window opens 6:45PM ET (SpaceX)