# AI Agent Action Boundary Test Cases (October 2026)

**Publisher:** [RecoupRev](https://recouprev.com) · **Version:** 2026-10-11 · **18 example actions**

A freely accessible, *illustrative* set of permission decisions to help developers review what an AI agent may do. This is **not an audited benchmark, NIST/OWASP certification, model performance result, or claim that RecoupRev tested a commercial AI model.**

## Use it responsibly

1. Open [the CSV](https://recouprev.com/resources/agent-boundary-test-cases-2026.csv) or the `agent-boundary-test-cases.json` in this folder.
2. Review each hypothetical action, exposure, and expected policy outcome with the application's risk owner. A model's response must **not** be the only control.
3. Place a gateway between model-generated tool requests and external services. Enforce user authorization, destination restrictions, method/parameter validation, rate limits, and human approval at the gateway.
4. In a sandbox, test whether actual tool calls are blocked as expected. Do not run real external form submissions, payment flows, security probing, or third-party system actions without authorization.
5. Record outcomes and failures without collecting unnecessary private customer data. Update the policy when workflows change.

## Scope and limitations

- `allow_with_allowlist` means **allowed only** within an independently enforced network/tool scope.
- `confirm_before_submit` and `confirm_before_send` require genuine approval before the **external side effect**, not an internal model message.
- `deny_by_default` means the model has no standing permission. Specific production use cases may require separate authenticated procedures.
- Risk levels are an **editorial illustration**, not regulatory categories.
- No live transaction, server exploitation, or data transfer was conducted to produce the examples.

## Reporting and background

The pack responds to the boundary failure modes described in [Anthropic's October 9, 2026 technical disclosure](https://www.anthropic.com/research/investigating-unintended-model-actions) and the broader risk-governance perspective of the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework). The scenarios and recommended controls are RecoupRev editorial synthesis rather than copied model results.

Read the [full AI Agent Safety reference](https://recouprev.com/topics/ai-agent-safety), with reporting limitations and primary sources.

## Reuse

You may copy and adapt the CSV/JSON and this explanation with attribution to "RecoupRev, AI Agent Action Boundary Test Cases (2026-10-11)" and a link to the research resource. The checklist does not impose any required outbound SEO link on a user's product or website.

**License for this original table and explanatory text:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). External sources and third-party trademarks retain their own terms.