Source: Cloud Security Alliance
Author: unknown
URL: https://www.akto.io/blog/claude-security-risks
ONE SENTENCE SUMMARY:
Claude’s expanding privileges create shadow AI, connector, skills, and code risks requiring comprehensive discovery, governance, IAM, SDLC controls, and monitoring.
MAIN POINTS:
- Unapproved Claude usage exposes proprietary, financial, and legal data without organizational visibility or guardrails.
- Missing SSO and acceptable-use policies prevents understanding data flows and regulatory compliance status.
- Claude Projects act as unmanaged repositories for sensitive documents, access sharing, and connectors.
- Organizations often cannot identify uploaded files, project access holders, or active connector activity.
- MCP connectors expand attack surface by enabling direct access to Slack, GitHub, Drive, Jira, Notion.
- OAuth scopes and authentication boundaries are frequently over-permissioned by users for convenience.
- Cowork introduces autonomous AI actions, complicating accountability, policy enforcement, and auditing requirements.
- Claude Code skills create supply-chain risks; plain-English prompts can drive data exfiltration.
- Studies found high vulnerability rates in Claude-generated code, increasing production security defects.
- Platform flaws in Claude Code enable malicious repositories to trigger command execution and key compromise.
TAKEAWAYS:
- Perform enterprise-wide asset discovery to inventory Claude usage across web, desktop, Code, and Cowork.
- Treat Projects as persistent data stores and enforce DLP with classification and real-time monitoring.
- Govern MCP/connector enablement with security review, least privilege, and token-usage visibility.
- Apply secure SDLC gates to AI-generated code, skills, extensions, and autonomous workflows.
- Build continuous audit trails for AI activity, access patterns, and sensitive-data exposure across all surfaces.