Suprvisr TLDR - July 13 2026

You spend fifteen minutes every day reading about what AI can do. Cumulus spends that time doing tasks autonomously for you. It's the AI teammate that works with your data (securely!): give it a goal and it runs the job end to end, pulls the data, drafts the report, and hands back finished work you can send as is. Every step logged, so you can always see what it did and what it touched.

No credit card required.

On July 8, OpenAI released GPT-Live, a voice model that listens and speaks at the same time. It murmurs "mhmm" while you're still talking and stops mid-sentence when you cut in. The awkward pauses that always gave away it was a machine are almost solved. More than 150 million people use ChatGPT's voice features every week, and for those people, voice now sounds much more natural. Look for "full duplex" voice models coming to all major AI players.

The next day, GPT-5.6 went public. We told you in June it had been locked up: announced, then handed to about twenty vetted organizations at the US government's request over its cybersecurity capability. The Commerce Department signed off twelve days later. OpenAI rates all three tiers "High" for cyber risk, and says none of them can autonomously run a complete attack campaign. GPT-5.6 comes in three flavours: Sol, Terra, and Luna, in order of performance (and of course price!).

Riding on top of it is ChatGPT Work, which takes a goal and hands back the finished spreadsheet or deck. The coding tool it grew out of has already spread past engineering, where a million people are using it for jobs that involve no code at all. That's where this is pointed: ordinary office work, done by something nobody hired. (Shameless plug for our Cumulus platform, which can do it too!)

The shift for you: none of this does anything until you hand it your inbox, your files, and your logins. That is a trade-off that should make you think twice, as AI providers are now focused on keeping the data you give them for extended periods of time just after a new launch.

That's the percentage of all internal hacking challenges that Sol by OpenAI was able to solve, and it is why the model spent twelve days waiting on a government sign-off. It is now available to anyone with a credit card. This is a striking pattern for new LLM releases, and was the reason Anthropic's Fable was pulled abruptly from the public.

🔓 An AI ran most of a ransomware attack by itself. Cloud security company Sysdig documented the first extortion operation where the intrusion was driven by an agent rather than a person at the keyboard. It hit a failed login, worked out why, and got in 31 seconds later. A human still chose the victim and set up the infrastructure, but the tactical work that used to require a skilled operator can now be rented. Many companies we work with are taking security and keys seriously, and key vaults exist for this reason. Ask your teams where your vulnerabilities are and see if you can stress test with open-source models (or ask Suprvisr AI to help!).

💰 The underdog is out-earning the household name. Anthropic, the maker of Claude, is now bringing in more revenue than OpenAI, roughly $47 billion a year against a reported $25 to $33 billion. OpenAI still has far more users, over a billion on ChatGPT. So who is "winning" now depends on whether you count sign-ups or dollars, and the gap is closing.

🔒 Google is training on your uploads unless you say no. Google changed a privacy setting in June, and TechCrunch flagged what it actually means to the public on July 6. If someone on your team photographs a document to search it, snaps a picture to identify something, or drops a file into Google Translate, that upload can now be used to train Google's AI. It's switched on by default, so nobody had to agree to anything. Google Photos is left alone, but the contract someone translated and the screenshot someone searched are not. There's an off switch buried in your Google account settings, and it takes about twenty seconds to find. This trend is consistent across all major AI companies, which are using their existing resources as AI training camps: Meta has Facebook, OpenAI uses Reddit, and Google uses... well, everything.

Trends are showing AI voice tech will keep getting better, and within a year, people calling your competitor won't be sure they reached a person or AI. Statistics are showing it won't stop them from making a purchase. Your business should start considering now: does an agent answer your phone in the future? If so, where does the recording and data go? How do you put bumper rails on those calls? Connect with Suprvisr AI if this is something you are curious about!