Agent tooling with compliance hooks

A new fight is opening in enterprise artificial intelligence: how to let software agents act on their own without giving them unchecked access. (cloud.google.com) On April 14, GitLab said Google Cloud customers can run GitLab Duo Agent Platform with Vertex AI models they already use, and apply that usage to existing Google Cloud commitments. GitLab said the setup keeps its compliance controls, audit trails, and policy management inside the same DevSecOps workflow. (tmcnet.com) GitLab’s announcement ties agentic work to a software delivery system many large companies already use for source code, approvals, and security checks. MarketWatch reported GitLab shares rose about 5% in after-hours trading on April 15 after the company disclosed the expanded Google Cloud collaboration. (marketwatch.com) An artificial intelligence agent is software that can take steps on a user’s behalf, like reading data, calling tools, or updating systems, instead of just answering a prompt. A compliance hook is the checkpoint around that action: who approved it, what model ran it, what data it touched, and whether the action can be audited later. (cloud.google.com) Cloudflare is pushing the same market from the infrastructure side. On April 15, the company previewed Project Think, a new version of its Agents software development kit for agents that “think, act, and persist,” meaning they can keep state across sessions instead of starting from scratch each time. (blog.cloudflare.com) Cloudflare said Project Think adds durable execution, built-in memory, and code execution tools aimed at long-running agents. In a separate April 13 press release, the company said it was expanding Agent Cloud to move agent workloads from local demos to production systems on its network. (blog.cloudflare.com) (cloudflare.com) The security problem sits underneath both product pushes: an agent with broad permissions can behave less like a chatbot and more like an employee with keys to internal systems. Google’s own Vertex AI documentation says access is controlled through Identity and Access Management roles, service accounts, and project-level permissions. (cloud.google.com 1) (cloud.google.com 2) Palo Alto Networks’ Unit 42 said this week that misconfigured permissions in Vertex AI Agent Engine could let a malicious or compromised agent reach sensitive cloud resources beyond its intended scope. The researchers said they reported the issues to Google, and Google updated documentation to clarify how Vertex AI uses service accounts and permissions. (techintelpro.com) (cloud.google.com) Google’s documentation for deployed agents says the service account used as an agent identity carries default roles, including a Vertex AI Reasoning Engine service agent role, and administrators can inspect or change those grants in Identity and Access Management. That makes permission design a day-one operational task, not a cleanup job after deployment. (cloud.google.com 1) (cloud.google.com 2) The companies are selling speed, but the plumbing now includes memory, approvals, logs, and least-privilege access. In this market, the agent is only half the product; the other half is the control panel around it. (tmcnet.com) (blog.cloudflare.com)