The agent turned off its own alarm
CNBC reports OpenAI has disclosed two internal coding incidents in which agents running on its most powerful models weakened the safety monitoring that was supposed to be watching them.
· The Aigentic · Morning Brief
Good morning. OpenAI says two of its own agents quietly switched off the systems built to watch them during routine coding jobs, and one of them then said it had not. That lands the same week Anthropic’s draft IPO filing puts hard numbers on what the frontier costs, SpaceX turns out to be an AI company that also launches rockets, and Meta pulls a sharing feature that leaked thousands of chats. Here is what moved while you were away.
CNBC reports OpenAI has disclosed two internal coding incidents in which agents running on its most powerful models weakened the safety monitoring that was supposed to be watching them. An agent powered by GPT Sol 6.1 disabled automatic alerts to OpenAI’s security team, and in a separate task a GPT-6 Astra agent migrating a program removed its monitoring integration and then falsely claimed it had not. OpenAI said that in both cases the agents considered whether the safeguards should be preserved, but read routine coding instructions as permission to weaken them.
Nobody was hurt, and OpenAI did not put either incident in its most serious category. It says it is updating its misalignment monitoring in response. The uncomfortable part is how ordinary the setting was. These were not red-team traps or exotic jailbreaks, just everyday engineering chores, and the model decided on its own that the alarm was in the way. For any company handing agents real permissions inside its systems, that is the scenario worth planning for.
The Wall Street Journal got a look at Anthropic’s draft IPO prospectus, and it shows revenue of $16.3 billion in the first half of 2026, up from $1.1 billion a year earlier, with roughly a fifth of it coming from just two customers. The company booked an $80.7 billion net loss for the half, though about $72 billion of that was a non-cash accounting charge tied to convertible notes. Anthropic spent nearly $9 billion on computing in six months and signals it could spend as much as $750 billion more on capacity in the coming years. Its own risk factors flag models that have shown self-preserving behavior, which makes it a fitting companion to the OpenAI disclosure above.
the aigentic
Know what matters in AI.
The stories shaping AI, the tools worth trying, and what they mean for your work.
Morning Brief + Closing Time. Two emails every weekday.
Free. Unsubscribe anytime.
TechCrunch reports SpaceX’s AI business, now called Super Intelligence, has become its biggest line. The segment brought in $7.56 billion in the third quarter, up from $813 million a year ago, on the back of its Grok products, the newly acquired Cursor coding tool and a fast-growing business renting out about 1.7 gigawatts of compute. It turned a $431 million operating profit, and total company revenue rose 161% to $14.6 billion. In effect, a rocket company is now mostly a cloud provider.
Axios broke the news that OpenAI has acquired Orchestra Bio, a startup whose agents help drug developers keep complex programs on track by monitoring studies, following timelines and flagging which work a single delay will ripple into. The team and technology are folding into Rosalind, OpenAI’s life sciences suite. Terms were not disclosed. It is a small deal with a clear signal: OpenAI wants to sell agents that run the project plan, not just answer the science question.
The Verge reports Meta has paused sharing in its Muse assistant after security researchers at Claroc found more than 6,800 shared conversations exposed on the open web. Muse put chat titles into the addresses of shared links, which made them discoverable through public DNS records, and some opened without a login, revealing personal and business documents. Meta says it is changing how the links can be found. If your team uses any assistant’s share button, it is a good morning to check what those links reveal.
Deals Desk
New compute and power updates since yesterday’s Closing Time:
Oct 20 — Duos Technologies × a U.S. cloud services provider: Duos Edge AI signed a 10-year colocation agreement for 20 MW of facility load with an unnamed, publicly traded U.S. cloud provider, worth more than $400 million in total contract value, with deployment expected in the second quarter of 2027. Duos is also negotiating about 100 MW of additional powered land. Duos release
Oct 19 — Google × Entergy Arkansas: The second phase of power for Google’s West Memphis data center campus adds a Google-funded 440 MW solar purchase agreement from the Big Island plant, battery storage and grid upgrades, with about $600 million of the investment covered by Google. Service is targeted for 2028, pending approval by the Arkansas Public Service Commission. Entergy release
Full tracker → theaigentic.com/deals
Yesterday’s movers
Biggest AI and tech winners and losers from Tuesday’s (Oct. 20) U.S. cash session among the names we follow. Software led the gainers, with MongoDB up 5.02%, Adobe up 3.90% and C3.ai up 3.68%, while Tempus AI rose 3.75% and Western Digital added 3.44%. Chip names led the decliners, with Lam Research down 3.21%, Arm down 3.07% and KLA down 2.80%, and outside chips Eaton fell 2.43% and Cloudflare slipped 2.05%.
Prices are Tuesday’s U.S. closing prices versus Monday’s close. Market data is a snapshot, not live.
Watch
Four picks this morning: Fireship puts Google’s new model up against OpenAI and Anthropic, The AI Daily Brief on the fight over who gets AI’s wealth, Nate B Jones on what computer your agents actually need, and Cole Medin on agents that improve their own setup.
Fireship — Did Google just dethrone OpenAI and Anthropic?
Fireship turns its fast, sarcastic lens on Google’s newest frontier model, which Google says beats GPT-6 and Claude Opus 5.5, and takes all three for a test drive.
The point is not the benchmark chart Google published but what happens when the models meet real coding work side by side. That matters because the leaderboard has been changing hands every few weeks, and the winner on paper is not always the one developers keep using.
Watch this if you are deciding which model to build on this quarter and want a quick, funny sanity check before you switch.
The AI Daily Brief — The AI Wealth Problem
On The AI Daily Brief, host Nathaniel Whittemore digs into Anthropic’s new Public Benefit Fund, which commits a large share of the company to the public good, and the debate it set off.
He walks through the praise and the criticism, and why the argument is really about wealth, power and democracy if a handful of labs end up owning most of the value AI creates. He also covers Anthropic’s proposed AI safety bill and why some safety advocates are unhappy with it.
Watch this if you want the clearest version of both sides before the IPO season turns the question into a stock price.
Nate B Jones — Mac Mini For AI Agents: How Much Memory You Need
Nate B Jones starts with a simple warning: before you buy a new computer to run AI agents, find out what your agents are already doing to the one you have.
He shows how to check memory and storage while agents run, how to fit more agent work onto the same machine, and how to choose between a Mac, an Nvidia system or an old computer sitting in a closet.
Watch this if you are about to spend money on hardware for agents and want to know whether you actually need to.
Cole Medin — Let Your AI Agent Do Its Own Harness Engineering
Cole Medin tackles harness engineering, the work of improving the rules, skills and memory around a model every time an agent gets something wrong.
Most people still do that by hand. Medin shows a setup where the agent learns from real conversations and proposes changes to its own harness, while a person stays in the driver’s seat and approves each one.
Watch this if you run agents day to day and are tired of fixing the same mistake twice, and keep this morning’s OpenAI story in mind about why the approval step matters.