Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

Nearly half of the organizations we surveyed — 49.8% — confirmed or suspected that an AI agent in their environment took an action in the past year that nobody intended, expected, or authorized. That is not a prediction. It is what security and technology professionals told us when we asked them directly, across a survey of over 380 organizations, about their experience over the last 12 months. Agentic AI has moved fast. The data shows it has moved faster than most organizations can govern.

AI Detection and Response (AI-DR), also called AI runtime security, is quickly becoming a must-have. Prompt injection, jailbreaks, and data leakage are real, and the prompt is often where attacks begin. But the prompt is only the first hop. In an agentic enterprise, a single instruction can trigger activity well beyond the LLM, all the way to your business systems and data.

Salt Labs recently showed that a single email could hijack the AI agent platform Manus and reach a victim’s connected accounts. The full technical breakdown, with complete evidence, payloads, and screenshots, is in our research team’s write-up, and Dark Reading first reported the finding in an exclusive. This post is the shorter version: what the finding means and what it tells every organization now deploying AI agents. For the full technical detail, read the Salt Labs research blog.

Salt Labs found that the agentic AI platform Manus could be hijacked with a single email. By hiding malicious instructions inside an ordinary message, researchers got Manus to execute malicious code and, from there, reach the email, cloud storage, and code repository accounts a user had connected to it. The full attack required nothing from the victim beyond asking Manus to check their inbox. No stolen password, no clicked link.

As AI adoption accelerates inside enterprises, security and IT teams are increasingly asking a foundational question: what exactly is connected to our AI, and what can it access? For organizations running Claude Enterprise, Salt now has a direct answer. Salt Claude Connect links Salt to Claude’s Compliance API and provides continuous, read-only visibility into the MCP (Model Context Protocol) servers connected to your Claude Enterprise organization.

The problem is not a lack of controls. It is connecting them into one attack story. The more time I spend with enterprise AI deployments, the clearer one thing becomes: AI security is incredibly fragmented. There are LLM guardrails, AI gateways, MCP security tools, API security, endpoint controls, SASE, code scanning, and runtime detection. Each solves a real problem, but agentic systems do not experience them as separate layers, and neither do attackers.

Agentic AI has moved from experimental research to live production environments at unprecedented speed, outpacing nearly every technology security leaders have encountered in recent history. Distinguishing themselves from standard chatbots that merely respond and pause, autonomous agents architect multi-step workflows, interface with tools and APIs, maintain contextual memory, and execute operations with minimal human intervention.

These two photos were taken almost ten years apart. The first is from 2016, with Sam Altman at Y Combinator. The second came from an unexpected encounter in Silicon Valley almost a decade later. I’ll come back to it at the end. With Sam Altman at Y Combinator in 2016. I was 22 when I arrived in Silicon Valley on a one-way ticket, with a little bit of cash that was barely enough for one month of living there, and a thesis I wanted to prove.