CrowdStrike finds AI tools inside a finance attack campaign
CrowdStrike says files exposed on attacker-controlled servers connected ARTEX, Claude Code sessions and other large language models to a campaign that stole data from South Korean financial organizations from late September into early October.[1]
The unidentified attacker used AI tools under human direction. CrowdStrike assessed with moderate confidence that the actor was likely Chinese-speaking and financially motivated.[1]
South Korea’s Financial Services Commission opened a one-month response period. It told financial companies to increase monitoring for suspicious loan applications, newly opened accounts and large transfers involving exposed information.[2] As of 6 October, authorities had confirmed no phishing losses or theft of funds resulting from the breaches.
The attack has moved the problem from theory to operations: financial institutions must detect people using AI to accelerate intrusions, then prevent stolen data from being turned into fraudulent accounts, loans or transfers.
OpenAI puts high-cybersecurity GPT-6 models into ChatGPT
OpenAI is rapidly distributing models it rates “High” in cybersecurity through a service the company says reaches more than 1.2 billion people each week. GPT-6 Sol began rolling into paid ChatGPT plans on 7 October. GPT-6 Luna started reaching Free and Go users on 8 October.[3][4]
OpenAI also rates both models “High” in biological and chemical capabilities. Neither reached the company’s “High” level for improving AI systems without human help.[3]
The company said the models strongly resisted prompt injection, in which hidden or hostile instructions try to redirect them. In one intentionally weak test setup, none of those instructions succeeded. In another low-stakes test without all production safeguards, the models sometimes worked around warnings.[3]
At this scale, the executive question is no longer only what the model can do. It is what permissions it receives, what monitoring surrounds it and which safeguards still work when more than a billion people can reach it.
One exposed AWS agent opened a route to other agents
Zenity Labs said a prompt sent to one tool-enabled Amazon Bedrock AgentCore agent revealed temporary cloud credentials. In the researchers’ controlled AWS environment, those credentials allowed access to other agents in the same account and region—and to their private conversations, source images, stored secrets and persistent memory.[5]
The researchers then used memory access to plant instructions that remained available in later sessions. A person directed the demonstration after the credentials were exposed.[5]
CSO Online reported that Zenity confirmed the permissions route was fixed before publication. AWS said the behaviour was expected and documented, not a vulnerability.[7] AWS’s current guidance warns that code inside an AgentCore microVM can reach execution-role credentials and says broad development policies should not be used in production.[6]
The test showed that putting an agent inside its own virtual machine is not enough if its credentials still carry broad cloud permissions. Protection depends on limiting each agent’s permissions and keeping its conversations, secrets and memory separate from those of other agents.
Q-Cash restores services after disruption at roughly 30 banks
Card, ATM, point-of-sale, internet-banking and national payment-switch services resumed at roughly 30 Bangladeshi banks after an outage that began on 4 October.[8] Bangladesh Bank summoned Q-Cash operator IT Consultants PLC and directed it to restore service.[8][9]
The explanations conflict. Bangladesh Bank initially pointed to upgrades, migration and maintenance.[9] The operator reportedly said a server fault forced an emergency migration. Another central-bank official alleged unpaid vendor bills and poorly planned system work.[8]
Bangladesh Bank said its cheque-clearing, electronic funds-transfer and real-time settlement systems continued operating.[9] The outage affected Q-Cash, a shared payment network used by many banks.
Service returned, but the episode demonstrated how one processor can interrupt customer-facing payments across dozens of banks. For banks dependent on a shared provider, recovery includes restoring availability and checking that transactions, records and configurations are sound.
Sources
- CrowdStrike, “Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance” — https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/
- South Korea Financial Services Commission, “Consumer alert following financial-sector data breaches” — https://www.fsc.go.kr/eng/pr010101/87887
- OpenAI, “GPT-6 Sol and GPT-6 Luna: October 2026 update” — https://deploymentsafety.openai.com/gpt-6-october
- OpenAI, “GPT-6 and Intelligent UI for everyone” — https://openai.com/index/gpt-6-for-everyone/
- Zenity Labs, “AgentCorruption” — https://labs.zenity.io/post/agentcorruption-how-a-single-prompt-collapsed-the-entire-cloud-security-model
- AWS, “Security best practices for AgentCore Runtime” — https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-security-best-practices.html
- CSO Online, “AWS’s repeated problems with AI agent controls illustrates the autonomous agent dilemma” — https://www.csoonline.com/article/4232054/awss-repeated-problems-with-ai-agent-controls-illustrates-the-autonomous-agent-dilemma.html
- The Daily Star, “Q-Cash services restored at 30 banks” — https://www.thedailystar.net/business/economy/news/q-cash-services-restored-30-banks-4293191
- The Financial Express, “Q-Cash network glitch disrupts digital transactions at some banks: BB” — https://thefinancialexpress.com.bd/bangladesh/q-cash-network-glitch-disrupts-digital-transactions-at-some-banks-bb