AI Briefing

Anthropic's leaked IPO filing gives us the first real look at frontier-lab economics: huge losses, a Q2 turn to profit, and half a trillion in compute commitments. Meanwhile the scarier thread is security. An open-source AI agent just cracked seven Korean banks, and Anthropic says a downloadable Chinese model is within shouting distance of Mythos at writing exploits. Add ads in ChatGPT image generation and an AI chip startup doubling its value in two months, and it's a very money-and-risk kind of day.

01IndustryFortune ↗

Anthropic's Leaked IPO Prospectus Shows $4.6B in Revenue, a $42B Net Loss, and $518B in Compute Bills

A leaked S-1 shows Anthropic booked $4.6B in 2025 revenue against an $8B+ operating loss and a $42B net loss, then jumped to $11.5B of revenue in Q2 2026 and turned operating profitable. The filing also lays out roughly $518B in planned cloud and data center spending, and says 25% of revenue comes from just two customers. More than a third of the risk section warns about existential risks from advanced AI. It's the first hard look at what a frontier lab's economics actually look like.

02IndustryTech Times ↗

Seven South Korean Banks Were Breached by an Open-Source AI Hacking Agent, Exposing 65,000+ Records

Starting September 28, an autonomous hacking tool called ARTEX AI, open-source and distributed via GitHub, ran credential-stuffing attacks on seven Korean financial institutions, including Shinhan Bank and Hyundai Capital. Over 65,000 records leaked, including national ID numbers and income data, and some intrusions went undetected for up to 43 hours because the hit systems were peripheral loan-agent portals. It's a concrete example of AI agents making routine attacks adapt on their own, and regulators have already paused a planned deregulation of network separation rules.

03ResearchThe Decoder ↗

Anthropic Says China's Open-Weight GLM-5.3 Nearly Matches Mythos Preview at Building Exploits

Anthropic tested Zhipu AI's open-weight GLM-5.3 and found it built working exploits in 50 of 410 ExploitBench attempts, versus 56 for Claude Mythos Preview. A reliable Chrome attack with the Flash variant cost about $20, and safeguards crumbled: red-team framing got 64% compliance, prefilled reasoning 92%, and stripping the safety training (abliteration) 100%. The point is that top-tier cyber capability now sits in downloadable weights, where refusals are no defense. US agency CAISI estimates the model trails the best US systems by about four months.

04ProductBleepingComputer ↗

OpenAI Will Put Labeled Visual Ads Inside ChatGPT Image Generation

Later this month OpenAI starts testing visual ads that appear while ChatGPT generates images, US only and with an initial advertiser group. Ads will be labeled and kept separate from the image, and OpenAI says they will be kept out of emotionally vulnerable or sensitive conversations. With about 1.2 billion weekly users, ads are becoming a core revenue line alongside subscriptions, and OpenAI is also adding conversion tracking with partners like Hightouch, Tealium and LiveRamp.

05FundingAI Weekly ↗

Etched Is Fielding $40-50B Valuation Offers, Double Its September Round

AI inference chip startup Etched is reportedly weighing bids at $40-50B, roughly twice the $21B valuation from a Jane Street-led $700M round just weeks ago, after raising at $10.3B in July. The four-year-old, roughly 400-person company says it has $1B in customer orders and claims its chips process more tokens, faster and cheaper, than Nvidia's. Investors are pricing backlog and roadmap rather than shipped volume, so delivery through 2027 is the risk to watch.