SG-T1 Agentic AI Risk Bounding and Accountability Worksheet
Verification required. This is a working template, not a regulator-approved form. Where it names an instrument or a date, the claim comes from secondary sources or from memory of the law and is a pointer to check. This is not legal advice.
How to use: This is a blank form. Copy it and fill in the empty cells and blanks for your system. To copy it, follow the "Suggest an edit to this page" link in the footer to reach its Markdown source.
Wiki controls: AI-CTRL-023, AI-CTRL-025, AI-CTRL-037, AI-CTRL-005. The structure follows the four dimensions reported for the IMDA framework [Reported]. Read the framework text and align the headings with it.
| Field | Entry |
|---|---|
| Agent ID, owner | |
| Objective and tasks | |
| Systems and data the agent can access | |
| Actions it can take (read, write, spend, send, delete) |
1. Risk assessment and bounding
| Item | Entry |
|---|---|
| Data sensitivity | |
| Scope of actions; irreversible actions | |
| Connections to other systems and agents | |
| Limits (spend, volume, time, scope) | |
| Failure modes and worst case |
2. Human accountability
| Item | Entry |
|---|---|
| Accountable person per agent | |
| Approval checkpoints (which actions need approval) | |
| Escalation route and time | |
| Kill switch and who may use it |
3. Technical controls
| Control | Entry |
|---|---|
| Least privilege and scoped credentials | |
| Sandboxing | |
| Input and output filtering | |
| Prompt-injection testing | |
| Monitoring and anomaly alerts | |
| Action logging | |
| Staged rollout (pilot, limited, general) |
4. End-user responsibility
| Item | Entry |
|---|---|
| What users are told about the agent's abilities and limits | |
| Training provided | |
| How users report problems |
Decision
Decision: __ Approver: _ Date: Next review: ___