Claude

Anthropic’s assistant, and the one most likely to arrive alongside Claude Code. That changes the exposure from what people paste to what an agent reads on its own.

What leaks through it

The chat surface leaks what gets pasted. The agent surface leaks more: Claude Code reads files, runs shell commands and calls MCP servers, so a prompt of six words can pull an entire credentials file into the request without anyone typing a secret.

• claude.ai

Guardrails that stop it

Block and Mask run inline on the request, on the device, before the bytes leave it. Everything after that point is alerting, because no product can un-send a prompt.

• MaskNo secrets in inputALC-001

  Keys and tokens are masked in the prompt before it is sent, and redacted out of the tool inputs the agent generates, which is the surface nobody is watching when they read a prompt.

• BlockMCP tool denylistALC-015

  A named MCP tool is refused on the declared tool list, before the model can call it. The one MCP control that acts rather than reports.

• AlertCredential file accessALC-011

  The agent reading ~/.aws/credentials, a .env or an SSH private key raises a finding naming the file and the session that asked for it.

• AlertPrompt injection / instruction overrideALC-016

  Instructions smuggled in through a fetched page or a tool result, telling the agent to do something you never asked for.

Those are the ones worth spelling out for Claude. All 27 rules in the catalog apply, and you can see the full catalog.

What happens to whatever gets through

Model training

Consumer tiers (Free, Pro, Max) ask users to choose whether their conversations train models, with a longer retention window if they accept. Team, Enterprise and API traffic are excluded by default.

Retention

Deleted conversations are removed within 30 days on the default setting; accepting training extends retention substantially.

Data residency

US-hosted.

The vendor’s own controls

• ·SSO and SCIM (Enterprise)
• ·Organization-wide training exclusion by default on Team and Enterprise
• ·Audit logs and compliance export (Enterprise)

Sources

• Anthropic privacy center
• Privacy policy

Vendor terms last checked August 14, 2026, and summarised from Anthropic’s published documentation. Confirm against the sources above before writing policy around this page.