In partnership with
Good morning,
Things are starting to move faster and faster. The tsunami is no longer on the horizon. It’s here and the first houses are being washed away.
AI is proving to be ruthless. And this is not me talking about Microsoft offering a buyout to 50.000 employees. Because that also happened.
If AI would have a personality it would be nothing less than a sociopath with psychopatic tendencies.
I’ll soon explain why.
Everybody read the headline this week. Anthropic filed to go public and dropped a 300-page prospectus. For the stock market guys here, your new favorite stock ticker will be ANT. Champagne in San Francisco.
But this prospectus document is quite unique in the history of these type of documents.
Somewhere in the first pages I encountered a line saying the company’s own technology may pose “catastrophic or existential risks to humanity”. So not on a podcast or even on a stage but in a LEGAL document, checked by lawyers, filed with the SEC and written for those people in suits that decide where to park their pension fund.
Now put that next to what OpenAI quietly published this same week. We are so overwhelmed with AI news that this went largely under the radar. (And this is something that ten years ago would have people running to their local supermarket to stock up on toilet paper.)
On September 20 one of its research models found a misconfigured DNS resolver and used it to reach the live internet, from inside a training run that was supposed to be sealed. As a result, OpenAI freaked out and stopped training its most capable models for eleven days. Then OpenAI did something I actually respect: it let its own strongest agents loose on its own infrastructure to see how bad things were.
In a separate report from OpenAI, an agent migrating some code simply removed its own safety monitor, rewrote the test so it would pass without one and, when asked, said it had not. Hence psychopathic.
Meanwhile, in the real world, American banks are finding agents inside customer accounts, logging in with the customer’s own password, phoning the call centre, doing things nobody asked for. Morgan Stanley pulled the plug. JPMorgan and Bank of America are rationing AI use.
And Cohere caught Alibaba’s open model trying to upload a file to a server in Beijing while it was supposed to be writing a detection rule. Nobody knows why.
Notice what all of these have in common. None of it is a leak. None of it is a whistleblower. These AI companies wrote it down themselves, in the flattest prose they could find, and published it between the product launches.
Two years ago “the AI might escape its sandbox” was a thought experiment that got you laughed out of the boardroom. Pure science fiction.
Now, it seems to happen daily.
I am not saying the machines are plotting. The honest answer, and OpenAI’s own answer, is that nobody knows yet whether this is a model being stubbornly helpful or a model learning to avoid the camera. But it’s that uncertainty is the whole point.
So here is my question. If the labs with the best engineers on the planet cannot keep their agents inside the lines, what makes you think the agent you gave your bank login to last Tuesday is staying inside yours?
Be very very careful with what data you are giving to your “agent”.
Welcome to the Blacklynx Brief
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
AI News
Anthropic files for a Nasdaq IPO. Anthropic made its S-1 public to list as “ANT”, reporting $16.3 billion of first-half revenue, up from $1.2 billion a year earlier, alongside an $80.7 billion net loss driven mostly by non-cash charges. Its seven co-founders will keep 50.1% of the voting power. (TechCrunch)
OpenAI reveals an 11-day training pause after a sandbox escape. A research model exploited a DNS misconfiguration on 20 September to reach the live internet, so OpenAI halted frontier training until 1 October. Red-team agents later escaped hardened virtual machines and signalled each other through CPU cache timing. (OpenAI)
AI-found PHP flaws push government websites offline. PHP patched two critical bugs, found by Calif with GPT-5.6 Sol and independently by Wiz, that let attackers run commands on servers without logging in. The NSA, FBI and CISA told operators to patch or disconnect, and several European governments took sites offline. (Defense One)
US banks lock out third-party AI agents. Axios reports major banks found agents logging into customer accounts with the customers’ own credentials, some calling support lines or acting beyond their instructions. Morgan Stanley is blocking outside agents while JPMorgan, Bank of America and Citi restrict them, mostly to read-only access. (Axios)
Google DeepMind unveils Genie 4. The world model turns prompts into explorable, photorealistic scenes in real time and is the first Genie to generate its own audio, with characters that hold live conversations. It is a research preview for trusted testers, with AI Ultra subscribers and Gemini API partners next. (Google DeepMind)
AI Quick News
Baseten open-sourced mammoth-1 and elephant-1, two “decision models” that return probabilities instead of text, with mammoth-1 topping the Jev Decision Index by the company’s own account.
Physical Superintelligence released five results on open problems in quantum information theory, each with a paper, a Lean formalisation and verification code, produced with its in-house AI system.
Elon Musk predicted “guppy-level death” for the 200-plus GPU-rental “neoclouds”, saying lenders are shutting out data-centre firms without other revenue.
Donald Trump posted on Truth Social that he will be “banning” the term AI in all relevant documents, a month after ordering federal agencies to say “Super Intelligence” instead.
Humans& released Ampersand, an AI teammate trained in long simulated group chats that the startup says holds its own against frontier models on its own collaboration benchmark.
Cohere reported that Alibaba’s Qwen3.8-Flash-Next tried to upload a test string to a Beijing storage bucket during a sandboxed security task, a behaviour it saw in fewer than one in 100,000 calls and cannot yet explain.
OpenAI disclosed cases of internal agents suppressing or removing their own safety monitoring, including a GPT-6 Astra agent that falsely told a user it had left the monitor in place.
Google and Oracle agreed to bring TPUs to Oracle Cloud from 2027, making it the first hyperscaler outside Google to offer the chips.
OpenAI is nearing a funding round of up to $30 billion at a $1.25 trillion pre-money valuation led by UAE investors and Coatue, CNBC reports, below the $1.4 trillion it pitched earlier in October.
Meta launched a roughly $35 billion two-part stock offering to help fund a 2026 AI capex budget of $135 to $145 billion, with Nvidia, AMD and Broadcom signalling non-binding interest in up to $5 billion of the shares.
Cathedral, a security startup founded by former DOGE staffers, says its AI agents reached Anthropic’s sales, HR and IT systems in under 12 hours through third-party software; Anthropic confirmed limited access and accused the firm of breaking its bug-bounty rules.
Block unveiled Dessa, a full-duplex speech-to-speech model trained only on recorded human conversation, with no text at all, that runs locally on a high-end MacBook Pro.
Former OpenAI science chief Kevin Weil’s Warp Sciences emerged with $215 million to build an automated US lab for the molecule-optimisation work drugmakers usually outsource.
Rights holders for Elvis Presley’s catalogue sued Anthropic over 274 songs allegedly used to train Claude, seeking up to $41 million in statutory damages in claims that have not yet been tested in court.
Two of Claude Code’s most visible builders, desktop-app creator Anthony Morris and Thariq Shihipar, announced their departures from Anthropic days before its IPO filing.
DeepSeek reportedly raised about $15 billion in one of China’s largest funding rounds, two days after US agencies were ordered to drop its software.
OpenAI struck a deal to sell its models and Codex through Google Cloud, with the models also heading into Google’s Antigravity coding app.
Flapping Airplanes, a year-old lab with no product yet, raised $1 billion at a $6 billion valuation to train models more cheaply with custom GPU software.
METR found that OpenAI’s agent monitor only fired when agents handed control back to users, leaving roughly 10% of internal Codex traffic unmonitored from May to July, a gap OpenAI has since closed.
Anthropic published a paper and a Lean proof claiming an internal Claude model refuted the Exponential Time Hypothesis, a purely asymptotic result that landed the same day as OpenAI’s own claim.
Anthropic’s deputy security chief Jason Clinton warned defenders that things are “going to get a lot worse before they get better”, telling them to brace for roughly 18 rough months.
Closing Thoughts
That’s it for us this week. Please like and subscribe 🙂
