Your midday AI news roundup for November 3. Today's headline act is Anthropic's warning: the company reportedly cautioned that AI is "very close" to the capability of "RSI" (Rapid Self-Improvement) — the point where AI can handle its own training. Meanwhile, OpenAI has brought on a former U.S. Secretary of the Army as a senior counselor for national security, stepping up its push to expand AI use in defense and cyber. On the consumer side, Apple posted record sales in China, and its earnings call also touched on revenue opportunities in third-party AI apps. Attacker-side AI is becoming real as well — we round up a case of massive charges from stolen API keys and a report of an autonomous attack tool confirmed in Japan.
Anthropic Warns It Is "Very Close" to RSI Capability
According to TechCrunch, Anthropic warned that AI is very close to the capability of "RSI" (Rapid Self-Improvement). RSI refers to the stage where AI becomes too good at training itself and the pace of its capability gains accelerates dramatically. For those concerned about catastrophic AI risk, it has long been the concept marking the branching point where a "catastrophe with no way back" begins.
It is no small thing that a party that develops models itself is stating that it is near that threshold. That said, the information gathered for this roundup doesn't include details about where or in what context the warning was made. Rather than taking it as definitive, we'll wait for official statements down the road.
OpenAI Hires Former Army Secretary Driscoll — Expanding AI Across Defense and Cybersecurity
According to Bloomberg, OpenAI has hired former U.S. Secretary of the Army Dan Driscoll. As senior counselor for national security and modernization, he will be responsible for driving the expansion of AI use in national and cyber security. Driscoll resigned as Army Secretary this past August after repeated clashes with Defense Secretary Hegseth.
It's a high-profile personnel move at a time when efforts continue to expand AI use across the military; the FT covered it the same day. It's a hire that shows AI companies and security institutions are steadily drawing closer together.
Apple Posts Record China Sales — Earnings Call Also Touches on Third-Party AI App Revenue
According to Bloomberg, Apple set a new sales record in the Chinese market in its September-quarter results. Greater China revenue climbed 21% to an all-time high, led by the flagship smartphone line. The Mac division also grew by double digits, and the services segment posted its best quarter ever in China. That said, results reportedly fell just short of the average analyst estimate. Amid headwinds from rising component costs such as memory and sluggish consumer demand, Apple still fared better than its competitors.
Around the same earnings news, TechCrunch reported another signal: on Monday night's earnings call, Apple mentioned the possibility of boosting revenue through "opportunities" with third parties. It appears to hint at a direction of turning third-party AI apps into a revenue source, though the specific mechanics haven't been disclosed. For now, there's nothing more than "we'll do something" — we'll wait for further details.
Attacker-Side AI Is Now Real — About $1 Million from Stolen API Keys, and Autonomous Attack Tool "ARTEX" Confirmed in Japan
Google Cloud's threat intelligence team reported a case in which stolen AI API keys resulted in roughly $1 million in usage charges at a single company. The scheme is known as "LLM jacking," in which third parties fraudulently use leaked cloud-based model credentials — leaving organizations with lax key management to bear the financial damage directly.
Attacker-side AI has been confirmed in Japan as well. On October 19, cybersecurity firm GMO Cybersecurity byLAC announced it had confirmed the use of "ARTEX," an autonomous AI penetration testing system, on infrastructure behind unauthorized access attacks on Japanese companies (as reported by ScanNetSecurity). As of early September, attackers "beginning to automate reconnaissance, intrusion, and detection evasion with AI agents" were still treated as an overseas phenomenon, but a little over a month later it has been observed on attack infrastructure in Japan as well.
The more convenient generative AI becomes, the more managing and monitoring API keys — and revisiting traditional intrusion detection — is turning into an urgent task for the defense side.
Alphabet in Talks to Acquire AI Assistant "Howie AI"
According to Axios, Alphabet is in talks to acquire Howie AI, an AI assistant startup. Google has been rebuilding its assistant experience around Gemini, and if the acquisition goes through, it is seen as a boost to that effort. However, this is a report at the negotiation stage, and the terms and outcome are unknown at this point.
OpenAI Introduces Git-Compatible Backend "Pando" — Built for Codex-Era Development Scale
On its official blog, OpenAI introduced "Pando," a Git-compatible backend that supports the company's own development. It's designed to handle the massive flow of development that Codex enables — an effort to scale while maintaining compatibility with the existing Git ecosystem. In an era when AI agents generate vast amounts of code, the same kind of limits will arrive for individual and enterprise development infrastructure too. In that sense, the efforts of the big players make for an interesting preview of what's to come.