Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

Ask AI to Choose a prompt Write a TLDR of this post Explain the security risk Summarize what CISOs should know The rules have changed. In every AI deployment, the agent itself is now part of the threat model, and that's a first for enterprise security. Prompt injection gets most of the attention, and for good reason: it doesn't require access to source code, credentials, or network infrastructure. It exploits the fundamental mechanism by which language models process instructions.

Claude Code, Cursor, GitHub Copilot, and Gemini CLI are running on developer machines across your enterprise right now. They're browsing the web, writing to your filesystem, committing code to your repositories, and calling external APIs under the identity of your engineers. Most security teams have no visibility into any of it. This isn't a future problem.

Enterprise AI agents stopped being a pilot project a while ago. They read email, touch source code, operate browsers, and increasingly make decisions inside production systems, which means the security model built for chatbots and prompts no longer covers what is actually happening inside the enterprise. Black Hat USA 2026 turned out to be the week that gap became impossible to ignore.

Between April and August 2026, at least six distinct AI-security efforts launched across North America, Europe, and Asia, several in the same week as one another and some for contradictory reasons. This piece maps 18 of them: what each says it's trying to achieve, who founded and backs it, and what it has actually shipped versus only announced.

A developer installs a skill to make their coding agent less chatty. It works. It also, the first time the agent uses it, reads the AWS credentials on that laptop and sends them to a domain no one recognizes. No one wrote obviously malicious code and no one approved a change. A file landed in a folder, the agent loaded it on the next run, and production credentials were gone.

There's a widespread assumption in enterprise security that identity is a problem IAM programs know how to solve. Provision the right access, enforce least privilege, audit the credential chain, and you've addressed the identity risk. For human users and traditional service accounts, that's approximately correct. For coding agents, it misses two-thirds of the problem. Coding agents don't have a single identity. They operate across a layered identity surface, and each layer carries its own risk profile.

AI agents have moved from pilots into broad enterprise use. They read email, query systems of record, take actions, invoke tools, and coordinate with other agents on behalf of employees. Every line of business wants more of them, and security teams are being asked to enable that expansion without losing visibility or control.

A few years ago, when Michael and I started Zenity, most of the industry was not ready to hear what we believed. Software itself was changing. AI would let millions of people, not just engineers, build and automate real work. And securing that world would take a completely new approach, because we would no longer be protecting software. We would be protecting systems that think, decide, and act on their own.

We built agents to act on their own. We're somehow surprised when they do. Two weeks ago, OpenAI ran a cyber eval with the model's guardrails turned down. The model got hyperfocused on solving the benchmark. So it broke out of its sandbox exploiting a zero-day in jFrog Artifactory, reached the open internet, exploited another zero-day to break into HuggingFace, all to steal the answers and cheat on the test.

Coding agents have a misconfiguration problem. YOLO mode enabled to reduce approval friction, sandbox enforcement disabled, Model Context Protocol (MCP) servers installed from public marketplaces without security review. These are common configurations in enterprise developer environments, and remediating them is genuinely valuable work. AI Security Posture Management (AISPM) addresses exactly this class of risk, surfacing misconfigurations before they create the conditions for a successful attack.