SentinelOne

I got to dig into one of the more unsettling incidents I've covered this year: an AI coding agent, running with --dangerously-skip-permissions, pip-installed a Trojaned LiteLLM package without a single human in the loop. The threat actor (TeamPCP) had compromised Trivy back on March 19th, stolen LiteLLM's maintainer credentials, and pushed a package with a data stealer, a fake "Sysmon" persistence mechanism, and Kubernetes lateral movement baked in. Here's the part I wanted people to see: SentinelOne’s Autonomous Security Intelligence never needed a signature for this package, and never needed to know LiteLLM was compromised. It watched the process tree, saw a Python interpreter spawn a subprocess to decode and execute obfuscated code, and killed it, all 424 related events, in 44 seconds. No analyst in the loop. AI coding agents with full permissions are a real, mostly unscoped attack surface right now. I walk through the whole chain in the new video, take a look: https://lnkd.in/gktyUSUK Thanks to everyone who helped in the process, including Adam Hall, Brian Goldfarb, Samantha Dondero, Caleb Tolin, Issa Zabaneh, Omri Misgav and others.

PHP2Shell is the name for two critical PHP flaws (CVE-2026-109618 and CVE-2026-109619, both CVSS 9.2). An unauthenticated attacker can execute code on the web server or read its process memory. The flaw is in PHP itself. WordPress, Drupal, Joomla, Magento, and Laravel are all exposed on an affected PHP version. Patching is the only complete fix. In the meantime, attackers with code execution still act on the host. They execute shell commands, drop webshells, download custom payloads, and open outbound connections. SentinelOne protects against this attack technique to stop such attacks at runtime. Read the full guidance: https://s1.ai/460199

Global adversary campaigns, region-specific TTPs, and hardening steps your team can act on, before attackers do.

Redact source code and customer data before it reaches an LLM

So Slack can keep saying yes to AI

Redact research IP and regulated data before it reaches an LLM