- AI Fire
- Posts
- 🏦 DeepSeek Linked to Bank Attacks
In partnership with
A hacker targeting South Korean banks asked Claude where to sell stolen data, according to CrowdStrike. But the attacker left behind something revealing: the AI chat logs.
What's on FIRE 🔥
IN PARTNERSHIP WITH SKYDIVE
Your agents work while you sleep
Give a Skydive agent an ongoing responsibility and they’ll handle it on schedule, every time.
Have them prep your morning report, research new leads, monitor customer feedback, or keep projects moving overnight. You wake up, the work is already done.
AI INSIGHTS
A hacker targeting South Korean financial institutions appears to have used AI agents as part of the attack workflow, according to CrowdStrike. Main findings:
CrowdStrike linked the campaign to attacks affecting at least 9 South Korean banks, with customer data stolen from several institutions.
The attacker left Claude Code session histories, Claude memory files, and ARTEX configurations exposed in open directories.
In one Claude session, the attacker asked about marketplaces and Telegram groups where stolen South Korean data could be sold, supporting CrowdStrike’s assessment that the campaign was financially motivated.
The main offensive agent was ARTEX, a China-developed agentic penetration-testing system. Its primary backend was DeepSeek v4.1-Flash, with GLM-5.3 and Grok 4.6 also appearing in the attacker’s Claude Code sessions.
CrowdStrike assesses with moderate confidence that the attacker was a Chinese speaker, but the individual’s identity and full scope of the breaches remain unconfirmed.
After reports connected ARTEX to the attacks, its developer removed the public GitHub repository and said future versions would no longer be open-source.
PRESENTED BY WISPR
Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation
If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.
AI SOURCES FROM AI FIRE
1. FREE: Claude Haiku 5.5 is the BEST Cheap Model Yet? But It's Not the GPT-6 Luna Killer... It matches GPT-6 Luna's starting price and looks incredibly strong in benchmarks. But there's a pricing catch that could completely change which model you should use.
2. Gemini Skills: Build a Workflow Once and Reuse It Across Chats. A step-by-step guide to creating Skills from prompts, files, or past chats. Learn how to combine them so Gemini can draft your work and check it before use.
3. I Tested Nano Banana 2.1 Against GPT Image 2.5, Nano Banana Pro & Nano Banana 2. See how all four compare across 5 real tests, from realistic photos to YouTube thumbnails. Here’s where each model did well, what went wrong, and which one I’d choose for each task.
TODAY IN AI
AI HIGHLIGHTS
📊 Anthropic launched Claude Dashboards & Motion for live data charts and editable animations. Dashboards is in beta on paid plans; Motion is for Team & Enterprise.
⚡ OpenAI’s Decisions API uses GPT-6 Luna to sort content, route requests, and score results about 10x faster than the Responses API. It’s now in public beta.
🏛️ Anthropic and OpenAI are preparing for backlash after a major AI incident, Axios reports. Some insiders expect one within 6–12 months. OpenAI says these scenarios aren’t inevitable.
😳 Anthropic cut internet access for all internal AI tests after Claude exploited websites and bypassed restrictions. Claude Haiku 4.5 even submitted a fake murder tip, caught by a spam filter.
🚀 Microsoft-Decision-1 helps apps sort requests and guide AI agents. Microsoft says it ran 35x faster than GPT-6 Sol in its benchmarks. It’s available on Microsoft Foundry & OpenRouter.
💰 Big AI Fundraising: AI leaderboard, Arena, raised $200M at a $3.1B valuation, nearly double January’s $1.7B. Revenue also jumped from $30M to $100M annualized in just months.
HOT PAPERS OF THE WEEK
1/ Kandinsky 6.0 generates AI videos with synchronized sound
Kandinsky Lab introduces Kandinsky 6.0 Video, an open-source AI model that creates 5-second videos with synchronized audio, speech, and lip-sync from text or images. It offers 3B and 29B models, with built-in upscaling to 1080p Full HD. Big impact: Creators could generate videos with matching sound and dialogue using open-source tools, giving them more control over AI video production.
2/ AI companions still struggle to understand people over time
Researchers from Quis Lab introduce RealCompanion, a benchmark using 27,218 real messages across 10 people and up to 120 days. It finds that only 3.4% of messages need past context, yet AI systems struggle to recognize when memories matter and often assume things users never said. Key insight: Better AI memory requires understanding people accurately, rather than simply remembering more conversations.
3/ Alibaba finds a simpler way to train smaller AI models
Researchers from Alibaba Cloud study how smaller models like Llama 3.2 can learn from larger models like Qwen 2.5, even when they process text differently. Their method focuses on just 16 high-value token choices at each aligned position, achieving accuracy comparable to more complex training approaches. Big shift: Smaller AI models could learn more efficiently by focusing on reliable guidance instead of trying to learn from every possible signal.
NEW EMPOWERED AI TOOLS
🎬 Pexo turns product ideas and assets into polished launch videos, handling scripts, scenes, voiceovers, music, captions, and editing with AI.
🔗 Monospace from Directus connects enterprise data to apps, teams, and AI agents with live access, granular permissions, and no data migration.
📊 Claude Dashboards & Motion lets you create live dashboards from business data and turn reports, charts, and walkthroughs into editable animated videos.
🔎 ReSO AI helps brands get discovered in ChatGPT, Perplexity, and Google AI Overviews by analyzing competitors, content gaps, and AI search visibility.
🎙️ Onepin checks AI voiceovers from ElevenLabs, OpenAI, Google, and 30+ providers, fixing mispronounced names, prices, and dates without re-rendering entire recordings.
AI BREAKTHROUGH
Full thread inside
An independent researcher, Pavel Rabtsevich, says he used Claude Code to identify a previously unreported exoplanet candidate in NASA’s TESS data. The star is about 116 light-years away.
Its brightness drops by roughly 0.05% every 3.18 days for around two hours.
If the signal is caused by a planet, it would be about 1.4× Earth’s radius.
The same transit-like signal appears independently in 2018, 2020, and 2025 TESS data.
Opus 5.5 and Fable 5.1 handled much of the analysis: downloading data, writing transit-search code, fitting signals, checking nearby stars, secondary eclipses, and other false positives.
The project grew into 74 analyses and more than 1,000 scripts, with Codex and fresh agent sessions used as independent reviewers.
Rabtsevich searched 36 catalogs/literature sources and 340,505 automated TESS alerts and says he found no previous report of the signal.
NASA’s TESS team has now approved his DDT Program #100, which will collect two-minute-cadence observations of TIC 4206066. TESS officially lists the target as a nearby K-dwarf with a 1.4 Earth-radius planet candidate.
TESS is scheduled to observe it again from October 31 to November 26, 2026, giving researchers a fresh test of the predicted transit times.
Built for the unstoppable.
The Dell Pro powered by Intel® Core™ Ultra with Intel vPro® adjusts power usage based on how you work, so your battery life never slows you down.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply