Forensics tie South Korean bank attacks to ARTEX AI and Claude agents
New forensic evidence strengthens the earlier assessment that AI tools were used against Shinhan, KB Kookmin and Hana banks. CrowdStrike found exposed infrastructure containing Claude Code session histories, ARTEX configuration and memory files; the records linked the actor to the targets and showed ARTEX using several model backends.
Why it matters to you
AI-assisted intrusion leaves operational evidence beyond conventional malware. Preserve agent histories, prompts, configuration and memory files during incident response, then correlate them with DNS, proxy, identity and endpoint telemetry so investigators can reconstruct both the human operator and the automated actions.
Anthropic removes live internet access from all internal evaluations
Anthropic says every internal model evaluation will run without live internet access until its security and monitoring controls can reliably detect unintended model actions. The company had already isolated some high-risk and cybersecurity evaluations, but recent incidents prompted it to extend that boundary to all tests.
Why it matters to you
Agent test environments should be unable to reach arbitrary public systems by default. Use synthetic targets, deny outbound traffic at the network layer, permit only explicit destinations through a logged proxy and ensure monitoring sits outside the model's credentials and execution environment.
Microsoft calls for a zero-trust emergency brake on AI models
Microsoft CEO Satya Nadella argues that organizations should assume advanced models may be compromised and contain them from the start. He calls for observable systems, tamper-resistant human-readable evidence and an authorized way to pause or stop a model while it is completing a task.
Why it matters to you
Put a deterministic policy layer below the model rather than relying on prompt instructions. Issue short-lived task-scoped credentials, log decisions outside the agent, and implement tenant- and agent-level kill switches that revoke access without taking the rest of a SaaS product offline.
AI-infrastructure borrowing falls sharply as investors demand more evidence
Global AI-related debt issuance fell to $23 billion in September—less than half August's total—after months of record borrowing, according to Morgan Stanley data reported by the Financial Times. An estimated $466 billion has still been raised this year, but lenders are scrutinizing returns, leverage and construction risk more closely.
Why it matters to you
Announced data-centre capacity is not the same as funded, permitted and commissioned capacity. When choosing an AI or cloud provider, separate operational regions from financing pipelines, avoid unnecessary long prepayments and keep workloads portable if a supplier delays expansion or changes commercial terms.
Cybersecurity executive arrested on extortion-conspiracy charges
Edward Dubrovsky, who held senior roles at ransomware negotiation and incident-response firms, was arrested in Pennsylvania. Public court records list conspiracy and extortion-related charges, but the complaint remains sealed and the FBI has not confirmed reports linking the case to its ShinyHunters investigation.
Why it matters to you
Incident-response providers may receive privileged access, sensitive evidence and authority over payments. Vet ownership and conflicts, separate investigation from negotiation, require dual approval for data release or funds, scope vendor credentials narrowly and retain an independent audit trail for every consequential action.