AI hacking tools hit South Korean banks; OpenAI’s revenue gap; Tencent’s bond plan

Compact Conversations for 2026-10-08: 5 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: AI-powered hacking tools breach multiple South Korean banks

A suspected single attacker used the open-source ARTEX tool, powered by AI models like DeepSeek and GLM-5.3, to breach several South Korean financial institutions, stealing over 25,000 customer records from Shinhan Bank.

Why it matters: Shows how AI-driven penetration testing tools can enable one person to conduct large‑scale attacks, raising concerns for financial security teams.

Source: The Decoder

Number to Know: OpenAI’s annualized revenue is $20 billion below prior signals

The Financial Times reports OpenAI’s annualized revenue is about $20 billion less than previously indicated, based on internal documents and interviews, suggesting emerging pricing pressures despite continued profitability.

Why it matters: A notable revenue gap that could affect OpenAI’s valuation and future research funding as the AI market matures.

Source: Reuters - AI

The Feed

Tencent considers $5 billion bond to fund AI ambitions

Tencent is exploring a five‑billion‑dollar offshore yuan bond to expand its AI GPU clusters and support the Hunyuan LLM series, indicating that even cash‑rich tech giants are using debt to accelerate AI infrastructure.

Why it matters: Highlights the scale of investment needed for AI and the shift toward debt financing among major tech firms.

Source: Reuters - AI

Claude adds animated video and live dashboard generation from text

Anthropic released two beta features for Claude: Motion creates animated explainer videos from prompts, and Dashboards turns data sources like BigQuery into live updating dashboards, both available across all plans including free.

Why it matters: Provides developers and analysts with quick ways to turn data and ideas into visual outputs without building separate pipelines.

Source: The Decoder

AI‑driven decline in call‑center hiring begins

Data from a major outsourcing firm shows a 15 % drop in new call‑center agent roles this year as AI handles routine inquiries, with cost savings up to 30 % on automated tasks.

Why it matters: Signals the real‑world impact of AI on employment in customer‑service roles and the need for workforce reskilling.

Source: Financial Times

One Thing to Try

Use the updated ttok tool (runs with uvx) to count tokens in text before sending it to an AI model, helping avoid surprise bills or truncated prompts. It includes a —list‑models command and works with scripts mixing code and prose.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: Today’s lead is a security story from South Korea. CrowdStrike reports that a suspected single attacker breached multiple financial institutions there, using AI-powered hacking tools. The tool is ARTEX, an open-source penetration testing tool that uses models like DeepSeek and GLM-5.3 to automate attacks. At Shinhan Bank, over 25,000 customer records were stolen. The attacker used ARTEX to automate reconnaissance and exploit chaining, reducing manual effort. CrowdStrike notes that the attacker attempted lateral movement, but detection tools blocked further progress. Host B: [thoughtful] CrowdStrike attributes this to a suspected Chinese-speaking attacker. The case shows how AI tools can let one person pull off breaches that used to require a team. ARTEX is on GitHub, so it’s widely available. The investigation is ongoing, and it raises questions for security teams. Security experts warn that as AI models improve, the barrier to launching large-scale breaches drops further, prompting calls for tighter monitoring of model usage in financial environments. This incident highlights how open-source AI-driven penetration testing tools can be repurposed for malicious use.

Host B: One number to know today is 20 billion dollars. The Financial Times reports that OpenAI’s annualized revenue is 20 billion dollars less than previously signaled. According to the FT, this figure comes from internal documents and interviews, not public filings. It’s a notable gap as the industry watches AI revenue figures. Analysts say the revenue shortfall could affect OpenAI’s valuation and its ability to fund future research, though the company remains profitable. The FT also notes that the gap emerged after a period of rapid enterprise adoption, suggesting that pricing pressures are beginning to show. Investors are watching closely for any updates in OpenAI’s upcoming financial disclosures.

Host A: Moving to the feed, Bloomberg reports via Reuters that Tencent is considering a 5 billion dollar bond sale to fund its AI ambitions. This is under consideration, so details may change, but it shows the scale of investment in AI by Chinese tech companies. Bloomberg reports that the bond would be denominated in offshore yuan and could fund expansion of Tencent Cloud’s AI GPU clusters, supporting its Hunyuan LLM series. Analysts see the move as a signal that even cash-rich tech giants are turning to debt to accelerate AI infrastructure builds. Host B: [with a small lift] Next, Anthropic shipped two new beta features for Claude. Dashboards turns data sources like BigQuery into live dashboards from text prompts, and Motion generates animated explainer videos from text and images. Docs, Slides, and Design now work across all plans, including free. Early Reddit users find Dashboards useful for quick data exploration. The Dashboards feature can pull live data every five minutes, enabling teams to track metrics without building a separate BI pipeline. Motion currently supports 1080p output and can animate slides, making it useful for quick explainer content. Host A: [conversational] And from the Financial Times, the long-predicted decline in call centre jobs has begun. The drop comes from reduced hiring, not firings. Data from a major outsourcing firm shows a 15 percent decrease in new agent roles this year, with AI handling more routine inquiries. The FT notes that the decline is most visible in outsourcing hubs like the Philippines and India, where AI handles routine inquiries such as password resets and order tracking. Companies report cost savings of up to 30% on routine handling, though they also invest in upskilling agents for more complex tasks.

Host B: One thing to try is Simon Willison’s updated ttok tool. It counts tokens using OpenAI’s tiktoken library and runs with uvx. He added a command to list models. [conversational] Use it to check token counts before sending text to an AI model to avoid surprise bills or truncated prompts. It’s based on OpenAI’s tokenizer, so it’s approximate for other providers, but good for planning. The tool also works with scripts that mix code and prose, giving a quick estimate of token usage before hitting API limits. Willison says the updated version fixes a Click warning and adds CI checks, making it more reliable for automation. Host A: For a quick test, try it on an email thread or a code file you’re about to paste into a chat. Running ttok on a typical README file shows roughly 250 tokens, helping you gauge context limits.

Host A: That’s Compact Conversations for Thursday. More AI news tomorrow. Until then, happy prompting.