Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

Claude is now a part of many employees’ everyday work. But even as a sanctioned application, it creates a visibility problem for security teams. Sensitive data can move into prompts, files, projects, and conversations, and teams need a practical way to see what happened and whether it violates policy.

A Russian-speaking threat actor known as “Trim” has spent the better part of 2026 systematically dismantling the guardrails on publicly available frontier AI models and rebuilding them as offensive tools. What started in March as a knowledge-sharing post on a Russian cybercrime forum detailing how to break Claude Opus into writing malware, had evolved by June into a fully productized, commercially marketed AI-powered penetration testing platform.

On July 14, 2026, attackers compromised release processes connected to the AsyncAPI open- source project and published malicious versions of four widely used npm packages with a combined reach of approximately 2.9 million weekly downloads.

In a controlled enterprise lab, we tested how far an agentic attack stack could go by harnessing a frontier model with an agent platform, MCP-enabled tooling, operational context, and enough autonomy to execute a complete attack path.

AI has moved from experimentation to operational reality. Last year saw a 50% rise in access to AI for employees, with 88% of organizations using AI in at least one business function. Competitive pressure is accelerating AI adoption. While this creates opportunity, it also creates a new level of operational risk.

Zero Trust was built to fix an older assumption: if you were inside the network, you were trusted. Then, Cloud, SaaS and remote work broke that, so security moved toward identity, device checks, MFA, least privilege, and continuous verification. But now, with agents, the messy bit starts after access. The agent reads a prompt, pulls context, chooses a tool, calls an API, and may trigger a workflow. The login tells you the agent is “trusted”.

Cato AI Labs has discovered two critical remote code execution (RCE) vulnerabilities in Cursor IDE, the popular development environment which, according to Cursor, is used by over half of the Fortune 500. Both RCE vulnerabilities, which we refer to as “DuneSlide,” achieved a 9.8 CVSS score, and involve breaking out of the IDE’s sandbox environment and were assigned CVE IDs CVE-2026-50548 and CVE-2026-50549.

AI did not create data exposure. It changed the consequences of it. Sensitive data, excessive permissions, and broad access policies have long existed across cloud environments. In the AI era, those issues are no longer passive governance concerns. They directly influence what AI users, copilots, agents, and applications can access, process, and expose. AI has turned a posture problem into operational risk. The challenge is not simply that more data is available.