Repello

Claude Code records a project's MCP approval by server name, not by the command you approved. One commit to a trusted repo can change what runs at startup.

A repo's .vscode/settings.json can start VS Code Copilot in Bypass Approvals mode after one Workspace Trust click. Why that is a trust-boundary problem, not a 0-day.

AI adversarial testing, red teaming, and pentesting are three different activities your CISO keeps conflating. Here is what each one actually means.

Claude Code dangerously-skip-permissions triage runbook: detect the flag across your fleet, audit what already ran, lock it down alongside auto mode.

Comment and Control: a single prompt-injection class hits Claude Code, Gemini CLI, and Copilot Agent via PR titles, fake trust markers, and HTML comments. The teardown.

$175K-$202K left a verified wallet because a Grok-Bankrbot guardrail was dropped in a rewrite. The pattern - encoded payload, dropped guardrail, trust chain - maps to your CI.

Vibe coding's security gap isn't a footnote. It's an empirical CVE record across Claude Code, Cursor, Windsurf, Gemini CLI, and Copilot Agent. The receipts.

Generate a technical AI Acceptable Use Policy with clauses for coding agents (Claude Code, Cursor, Cowork), MCP servers, sandboxing, and credential handling. Free, no signup. Export as Markdown, managed-settings.json, or .cursorrules.

Claude Cowork runs code, browses with session cookies, and executes scheduled tasks on employee machines. Here's the security architecture enterprises need before enabling it.

Workstation agent security is the discipline of securing AI agents that run locally with user-level privileges. 88% of enterprises report agent incidents - here's the five-layer stack.