Varonis is betting that AI coding agents and file-editing agents are the next unsecured perimeter, and it’s moving to own that surface before most security teams have mapped it. The company’s Atlas platform now covers the full Anthropic Claude enterprise stack, including Claude Code and Claude Cowork, adding runtime guardrails (real-time traffic cops that block bad prompts before they reach code or files), activity monitoring, and AI security posture management. The Claude Code and Cowork coverage<\/a> follows Varonis’s earlier Claude Compliance API integration and is available today.<\/p>\n The question this announcement forces is whether your current security stack has any visibility into what your AI agents are doing at the moment they act, not in the log review two hours later. If your developers are running Claude Code against production repositories, or knowledge workers are letting Claude Cowork rewrite files on local drives, the exposure window opens at access, not at breach. Organizations already deep in Anthropic’s ecosystem face this question now; those still evaluating agentic AI tools have a narrow window to set the governance architecture before agents are load-bearing.<\/p>\n The specific capability worth examining here is AI Security Posture Management, or AI-SPM. Traditional security posture management scans for misconfigured cloud buckets or overprivileged service accounts. AI-SPM extends that logic to agent-specific attack surfaces: memory files that persist instructions across sessions, Model Context Protocol configurations that can silently redirect an agent’s behavior, and skills that may exfiltrate data without triggering conventional data loss prevention rules. These aren’t theoretical risks. They’re the natural consequence of giving an AI agent persistent memory and file-system access, and most enterprise security teams haven’t written policies for them yet.<\/p>\n Varonis’s market signal here, with VRNS gaining 8.2% on the announcement day against a prior pattern of negative reactions to its AI-tagged news, suggests investors read this as genuine product differentiation rather than another partnership press release. The falsification condition is straightforward: if Anthropic or a competitor like Microsoft builds equivalent monitoring natively into the agent runtime and bundles it at zero marginal cost, the addressable market for a standalone Atlas shrinks fast. Until that happens, the CISO who has already approved Claude for developer use but hasn’t audited the agent’s data access permissions owns the most immediate decision this story creates.<\/p>\nWhat this means for your business<\/h2>\n