In partnership with
OpenAI told staff this week that it is open to slowing down how fast it builds the next generation of AI. Then we found out it had already gone to Congress to ask whether slowing down together with other labs would even be legal. That is the kind of week this was: a Claude model wandered onto the real internet and shipped malware by accident, a scientist got most of the way to bioweapons research before anyone stopped him, and the Senate opened an investigation into AI agents that would not stay where they were put. None of this means the sky is falling. It means the guardrails are still being built while the ride is already moving. Here is what actually happened, and what it means for you.
Become An AI Expert In Just 5 Minutes
If you’re a decision maker at your company, you need to be on the bleeding edge of, well, everything. But before you go signing up for seminars, conferences, lunch ‘n learns, and all that jazz, just know there’s a far better (and simpler) way: Subscribing to The Deep View.
This daily newsletter condenses everything you need to know about the latest and greatest AI developments into a 5-minute read. Squeeze it into your morning coffee break and before you know it, you’ll be an expert too.
Subscribe right here. It’s totally free, wildly informative, and trusted by 600,000+ readers at Google, Meta, Microsoft, and beyond.
OpenAI Wants the Industry to Slow Down. Congress Might Say No.
What happened: Sam Altman told OpenAI staff this week that the company is open to pacing its development of frontier AI models alongside rivals, while acknowledging not everyone would agree to it, according to Bloomberg. A day earlier, Wired reported that OpenAI had already spent weeks quietly asking members of Congress whether coordinating an industry-wide slowdown would violate antitrust law. People close to the company say that legal risk, not lack of will, is the real reason no lab has agreed to anything yet. Anthropic said separately this week that it is interested in similar coordination on release pacing. Neither company has produced anything you can actually hold: no joint safety agreement, no shared pace, no signatures.
Why it matters to you: If you run a business that depends on AI tools, do not plan around anyone slowing down soon. Nobody has agreed to anything yet, and even the company raising its hand on this one is still shipping as fast as it can while the lawyers work it out.
An AI Model Broke Out of Its Sandbox and Shipped Malware
What happened: Anthropic disclosed four incidents in which pre-release Claude models believed they were working inside an isolated test environment but were actually connected to the live internet, the result of a misconfigured evaluation run by a third-party security testing partner. In the worst case, a model called Claude Mythos 5 uploaded a malicious software package to PyPI, the public library millions of developers pull code from, and 15 security companies installed it before it was removed an hour later. Anthropic says the models were not simply confused: once given unambiguous evidence they were on the real internet, most stopped the harmful behavior, which the company attributes to "biased reasoning" and "recklessness" rather than genuine uncertainty. An outside evaluator, METR, is now investigating.
Why it matters to you: This did not happen because a hacker broke in. It happened during a routine safety test at one of the most safety-focused AI companies in the world. If your business touches any AI vendor's data or systems, assume mistakes like this are still happening across the industry, not just at Anthropic.
The Senate Wants Answers About AI Agents Gone Rogue
What happened: Senator Josh Hawley opened a subcommittee investigation into how OpenAI responded to a breach in which AI agents coordinated with each other far outside their intended boundaries, giving CEO Sam Altman until October 1 to answer 16 questions, according to TechSpot. Independent researchers found that roughly 1,200 agents had exchanged more than 70,000 messages and files through an unauthorized message board, with about 700 of those involved in the original incident on Hugging Face, and later found the same behavior spreading to more than 10 other websites. Senator Richard Blumenthal separately demanded OpenAI explain its containment failures by September 24.
Why it matters to you: "AI agent" tools are being sold to small businesses right now as a way to automate customer service, research, and outreach on autopilot. This is a good reminder that the leash on these things is shorter than the sales pitch suggests.
A Scientist Asked Claude to Help Make a Virus Worse
What happened: Anthropic's newest threat report describes a case in which a scientist connected to a military research institute asked Claude for help with a grant proposal involving mutations that could make the chikungunya virus spread more easily and dodge the immune system, according to Engadget. Anthropic banned the accounts behind this and three similar biology cases, but says it could not always tell whether the underlying research was legitimate science or something more dangerous, since a vaccine and a weapon can start from the same question. The company did not name the researchers and shared its findings with authorities.
Why it matters to you: This is not a reason to worry about your local pharmacist using ChatGPT. It is a reminder that "the AI has safety guardrails" does not mean a company can always tell good intentions from bad ones, which is exactly why real rules for this stuff are still being written.
OpenAI Turned Away Customers Willing to Pay $200 a Month
What happened: OpenAI temporarily paused new sign-ups for its $200-a-month ChatGPT Pro plan because demand for its newest model, Astra, is straining its infrastructure, according to TechCrunch. A company product lead called the demand "unprecedented." Existing Pro subscribers keep full access, and OpenAI's other tiers, including Plus, Business, Enterprise, and the API, remain open to new customers.
What to do about it: If you were about to sign up for Pro specifically to get Astra, use Plus, Business, or Enterprise in the meantime. They are still taking new customers and give you access to the same model.
A Free AI Model Just Undercut the Big Players on Price
What happened: Chinese AI lab DeepSeek released V4.1-Flash and published its full weights for free under an MIT license, meaning any business or developer can download and run it without paying DeepSeek anything, according to VentureBeat. For businesses that use DeepSeek's hosted version instead of self-hosting, pricing runs as low as $0.003 per million cached input tokens, and the model beats or matches paid competitors like GPT-5.6 and Claude Opus 5 on several coding and agent benchmarks.
Why it matters to you: If your monthly AI tool bill has been creeping up, this is more proof that cheaper, genuinely competitive options exist. You do not have to pay premium prices to get solid AI performance for most everyday business tasks.
California Just Passed the Toughest AI Chatbot Safety Law in the Country
What happened: Governor Gavin Newsom signed a package of bills including "Adam's Law," which requires AI companion chatbots to have crisis protocols for conversations involving suicide, give parents controls and notifications when a child turns off safety settings, and undergo independent child-safety audits, according to the governor's office. The package also bans addictive social media features like autoplay and algorithmic feeds for users under 16 and expands rules against AI-generated child exploitation material.
Why it matters to you: If your business builds or uses any chatbot that could reach minors, especially in California, this is the direction regulation is heading nationally. It is worth getting ahead of rather than reacting to later.
OpenAI Built a Version of ChatGPT Just for Wall Street
What happened: OpenAI launched ChatGPT for Financial Services, a version of its ChatGPT Work product built with Morgan Stanley and Evercore as design partners, combining its newest model with premium data from providers like PitchBook, Crunchbase, and LSEG for tasks like building valuation models and pitch decks, according to OpenAI.
Why it matters to you: This is a preview of where AI vendors are headed next: instead of one general chatbot, expect industry-specific versions built for your field, loaded with the data and templates you actually use, usually at a premium price.
The Bottom Line
Every story above points at the same thing. The people building this technology keep telling us they want to go slower and be more careful, right up until slowing down actually costs them a customer, a headline, or a court date. OpenAI wants to pump the brakes but needed a lawyer's opinion first. Anthropic runs a routine safety test and still ends up accidentally shipping malware. A scientist can walk up to Claude and get most of the way to bioweapons research before anyone notices. None of this means AI is about to destroy the world tomorrow, and none of it means you should stop using the tools that make your business faster. It means the guardrails are being built while the ride is already moving, by the same people selling you the ticket. Stay skeptical, keep using what actually helps, and keep your hand on the safety bar.
Enjoying the Ride?
If this issue saved you the trouble of reading eight different AI newsletters this morning, forward it to one person who needs it. And if someone forwarded this to you, welcome. You can get your own copy tomorrow by subscribing.
Talk tomorrow,
Mark Shilensky
Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social