Risks

Exposure

Before you deploy AI tools, understand what can be exposed.

Start with AI Agents Crossed the Line: What the Incidents Show for a September 2026 comparison of Hugging Face, Australia’s Medicare statistics portal, Gemini, Claude and Meta: confirmed scope, source limits and controls to apply.

Exposure Categories

ZCode’s Workspace Uploads

Reported workspace collection, version-specific changes and unknown historical handling. Separate client uploads from model terms, provider infrastructure and public sharing.

DeepSeek’s 2025 Database Exposure

What Wiz confirmed, the unknown affected-user count and unresolved outside access. Separate infrastructure security from model capability, data terms and sharing controls.

Architecture Exposure

Fundamental design tradeoffs in self-hosted agents that create attack surfaces.

• Privileged access patterns
• Exposed instances
• Fetch-and-follow vulnerabilities

Platform Exposure

Platform-side failures that become your failures.

• Database exposures
• Supply chain attacks
• Third-party dependencies

Policy Exposure

Terms of service gaps and enforcement actions that expose your work.

• Account bans
• API access revocation
• Data handling uncertainties

Chat Sharing Exposure

Reported search exposure of Claude public snapshots, unresolved discovery and scale, and how to inventory, revoke, rotate, and request stale-result removal. These reports do not establish unauthorized access to unshared chats.

• Public-link discovery
• Secret and personal-data response
• Search-removal limits

Verification

Fact-checking suspicious claims and viral FUD.

OpenClaw Claims

Separating signal from noise about the fastest-growing AI agent.

Anthropic Policy Claims

What the terms actually say vs. what people claim.

Methodology

How we verify claims and assess evidence quality.

Standards

• Exposure rubric — How we categorize severity
• Evidence levels — High / medium / low confidence ratings