- A verdict on governed prompts and tool calls before Claude runs them, from claude.ai, Cowork, Claude Code and Claude Tag in Slack
- Checks for prompt-injection and jailbreak attempts, limits on which tools an agent may call, approved models only, and 57 sensitive-data patterns
- A critical alert when the checks are switched off or a governance setting is weakened; 30 kinds of risky change or hook event in all
- Claude Enterprise activity every 15 minutes, counted as evidence towards 24 controls across 20 compliance frameworks
MELBOURNE, Australia, 29 October 2026: Kapālins, an Australian AI security company, brought guardrail integrity to Claude Enterprise. Its integration stops risky prompts and agent actions before Claude acts, and raises an alert if anyone switches those checks off.
Guardrails only protect while they stay switched on. Claude Enterprise, Anthropic’s AI platform for organisations, runs in the cloud, so a corporate network gateway cannot sit in front of it, and one change in its admin settings can quietly remove a check. Kapālins’ view is that a security team needs the check, and proof that the check is still there.
Stop it before Claude acts
Through Claude Enterprise’s Inference hooks, a beta feature, each governed prompt from claude.ai, Cowork, Claude Code and Claude Tag in Slack reaches Kapālins before Claude runs it, as do tool calls when tool-call checks are on. Kapālins answers allow or deny with the same Policy Shield rules as its AI gateway: checks for prompt-injection and jailbreak attempts, limits on which tools an agent may call, approved models only, and 57 sensitive-data patterns. It answers within 2.5 seconds by default; if it cannot decide in time, the organisation’s own setting blocks the prompt or lets it run unchecked. A person whose prompt is denied sees the reason.
Guard the guardrails
Through Claude’s Compliance API, Kapālins reads Claude Enterprise activity every 15 minutes. It raises a critical alert when someone deletes or changes the Inference hooks configuration, when a request goes ahead without a verdict, or when single sign-on, data residency or Claude Code’s zero data retention is turned off or weakened. Giving Claude’s agents more autonomy or reach, such as auto mode or a new MCP server, raises a high alert. In all, 30 kinds of risky change are watched.
One integration, three jobs
One record
Alerts join the same incident queue as the other AI Kapālins governs. With content capture on (it is off by default), chats, text uploads and Cowork and Claude Code sessions are checked against the same policies. Each denial carries a Kapālins reference ID that Claude records on its own denial activity, so the two records can be followed in either direction. The verdicts and activity count as evidence towards 24 controls across 20 compliance frameworks, including NIST AI RMF, the EU AI Act, MAS guidelines and APRA CPS 234.
Kapālins built the endpoint to Claude Enterprise’s published Inference hooks contract. It passed all 17 required checks, and in testing returned 95% of verdicts within 229 milliseconds.
“Integrity is about whether AI still behaves the way it was approved to,” said Jai A Singla, Chief Technology Officer of Kapālins. “For Claude Enterprise, that starts with the guardrails themselves. Kapālins stops the risky prompt and the risky agent action before Claude acts, and if anyone weakens those controls, the security team knows at the next check.”
Availability
The integration is available now to Claude Enterprise organisations. Inline verdicts carry no extra usage charge, and verdicts store no prompt text. Kapālins will be listed in Claude’s documentation as a Claude Enterprise integration in the coming weeks.
The integration is for Claude Enterprise; Claude Platform organisations provide activity events only. Kapālins returns allow or deny; it does not rewrite prompts or change how Claude responds.
The integration page, with a product film, is at kapalins.com/integrations/claude-enterprise.
About Kapālins
Kapālins is an Australian AI security and governance platform. It finds the AI an organisation runs, from approved copilots to tools nobody signed off, applies policy to prompts, data and agent actions, and turns compliance into evidence across 42 frameworks in 15 jurisdictions plus global standards. Kapalins Pty Ltd is based in Melbourne. More at kapalins.com.
