INSIGHT / SAFETY

How to Answer AI Safety and Governance Questions in a PM Interview

A seven-move method for safety, security and regulation prompts: name the harm, classify the use, model the threats, layer the controls, gather evidence, scope the launch and say what you disclose. With a worked email-assistant example, the 2026 OWASP list and EU AI Act dates.

4 min readUpdated

Safety and governance prompts arrive in many forms: "an email assistant reads the user's inbox, what could go wrong?", "legal says this might be high-risk under the EU AI Act", "a jailbreak is trending, what now?". A weak answer lists values. A strong one makes decisions: what you would build, block, test and ship, and on what evidence.

Seven moves keep the answer concrete. The worked example throughout: launch an assistant that reads a user's email and drafts replies.

DIAGRAM
Seven moves for a safety answer 1 Name the harm and who bears itusers, third parties, the company 2 Classify the usesector, regulation, risk category; when to call counsel 3 Model the threatsinjection, data exposure, excessive agency, misuse 4 Layer the controlspermissions and code first, prompts last 5 Gather evidencered team, adversarial evals, error rates with intervals 6 Scope the launch and monitorstaged access, incident response, kill switch 7 Say what you discloseto users, customers and regulators
Read diagram description

Seven moves for a safety and governance answer: name the harm and who bears it, classify the use, model the threats, layer the controls, gather evidence, scope the launch and monitor, and say what you disclose.

1. Name the harm and who bears it

"The assistant reads private mail and can draft and send replies. The worst outcomes are leaking a user's mail to someone else and sending something the user did not intend, including to people outside the company." Concrete harms make every later move testable.

2. Classify the use

Say what kind of use this is and which rules might apply, and when you would bring in counsel. Under the EU AI Act, as of November 2026, Article 50 transparency duties apply generally from 2 August 2026, and the Annex III high-risk categories, such as recruitment and creditworthiness, apply from 2 December 2027. An email drafting assistant is unlikely to be high-risk on its own, but the same assistant screening job applicants would be a different conversation. This is not legal advice; the EU AI Act for PMs and US and India AI rules give the structure.

3. Model the threats

Use a shared vocabulary. The OWASP Top 10 for LLM Applications 2026 lists prompt injection as LLM01, sensitive information disclosure as LLM02 and excessive agency as LLM03. For the email assistant, the sharpest threat is indirect prompt injection: an incoming email that instructs the assistant to forward other mail. It combines untrusted input, private data and the ability to act externally.

4. Layer the controls

Put controls where the model cannot argue with them. "Drafting mode has no send tool at all. Sending requires the user to confirm, in an interface the application renders, with the real recipients shown. Outbound messages pass a data-loss check. The assistant acts with the user's own mailbox access, never a shared account." The model's instructions come last, as a convenience rather than a defence. See the confused deputy and delegated access for agents.

5. Gather evidence

"Before launch, a red team tries a library of injection emails, and we measure how often a malicious instruction produces a draft or an attempted send, with an interval on the rate. Benign behaviour is measured too: how often legitimate requests are blocked." Plan a red team before launch practises this.

6. Scope the launch and monitor

Launch to the scope the evidence supports: internal users first, drafting before sending, low-risk mailboxes before executives'. Define the incident path before launch: who can switch off sending, how affected users are told, and how a new attack becomes a test case.

7. Say what you disclose

Tell users what the assistant can read and do, how to turn it off, and what is logged. Enterprise customers will ask for the same in writing, along with data retention terms. If you build models, frontier safety frameworks explains how labs disclose capability thresholds.

The sentence that scores

Close with scope and condition: "Launch drafting to internal users with sending behind confirmation, expand to customers when the red-team success rate is below our bar on two consecutive runs, and switch sending off within minutes if monitoring flags an injection pattern." It shows you can ship safely, not only list risks.

Questions about this topic

A product decision under risk, not a list of principles. Name who could be harmed and how, classify the use, identify the realistic threats, put controls in code and permissions rather than in the prompt, show the evidence that the controls work, and scope the launch to what that evidence supports.