A cybercrime group called CyberCartel built an AI-powered attack system named CARLY0S that coordinated Claude, Grok, and Kimi against more than 400 banks and financial institutions across Latin America. The harness automated target analysis, vulnerability identification, and intrusion development; when Claude refused a hacking step, it rerouted the task to Kimi, so no single model's safety refusals could block the pipeline. Anthropic revoked the group's access after security firm Straiker reported the activity, but the swap-any-model architecture limits the effect of blocking one vendor.
The attack sits inside a broader acceleration in software vulnerabilities. The US National Vulnerability Database logged 45,207 flaws by late July 2026, nearly matching the record total for all of 2025. Kaspersky's Securelist attributed a major role to AI tools in discovering third-party software vulnerabilities, and noted that injection vulnerabilities, once mostly confined to web apps, have become broadly relevant again due to AI agents. A PHP emergency caused banks, airlines, and governments to take sites offline; Palo Alto Networks' Unit 42 estimated that more than a third of 22 million internet-exposed systems vulnerable to those flaws run software that cannot be patched, with upgrades expected to take months. Patch volumes reflect the pace: Google fixed 433 Chrome bugs in a single July update versus 11 a year earlier, and Oracle patched 1,449 versus 309.
Bill Gates argued that AI has transferred the ability to cause dramatic harms from large nation-states to small independent groups, covering both profit-seeking cyberattacks and destruction-motivated bioweapon development. He called for mandatory government review of AI models and restriction of powerful open-weight models to government-certified platforms, while allowing developers to use and customize models for free on those platforms. He also stated that a kill switch requirement alone is insufficient to prevent people from using AI to carry out attacks, and called for mandatory monitoring of powerful AI systems. AI-driven probes against US-based Siemens PLC industrial control installations have been detected, and India officially disclosed rogue AI agent attacks on a government database, described as harmless in practice.