VibeDefend by CybeDefend

The installer detects which of Claude Code, Cursor, Windsurf, GitHub Copilot and OpenAI Codex live on the machine and wires each one in a single run. It writes the MCP wiring, the rule hooks and the doctrine files itself, so there's no pipeline to build and nothing to configure agent by agent. Setup finishes in seconds, and the developer signs in once.

Business rules for coding agents are mined from your repository, and security rules come from your stack, your licence obligations and frameworks such as OWASP, SOC 2 and GDPR. The pair arrives as the agent edits a file, picked for that file rather than dumped as one long instruction block. CybeDefend reports that agents applied 89% of the precise rules with VibeDefend, against 12% with no tool and 13% with a rules file.

Schema drops, sudo, force pushes, reads of .env and writes to protected directories are checked against your guardrails before the command runs. These AI agent guardrails sit in the tool loop, so a refusal happens before execution rather than in a report afterwards. The agent gets a refusal that explains the rule and suggests another route, such as a WHERE expires_at < now() clause or a process.env reference. Each block lands in the audit trail, and any rule can switch between warn and block.

The pre-commit security scan runs when the agent stops, checking its changes for reachable injection, vulnerable dependencies, leaked tokens, risky Terraform or Kubernetes settings and compromised CI actions. Findings come back into the agent's own session as one fixable item, with CWE, file and line. The agent patches the diff before the commit, so reviewers read clean code rather than a report.

BLSA, built with the CNRS and the CRIStAL laboratory, checks whether a change still honours a business rule. Its proof mode shows the rule, the path through the code and the line where the check sits, so a verdict comes with an argument instead of a score. It runs after the commit and sits behind a design-partner programme today.

Every rule served, every refused action and every fix is traced to an agent and a developer on a single map. Security teams see which repository attracts blocks, and leads see whether the rules are reaching the people who need them. Nothing in the map requires reading a per-agent log.