AI Builders Global Conference 2026
Oct 14, 2026 – Oct 16, 2026 · Virtual · scheduled
At this event: Salesforce
Ishween Kaur
Senior Software Engineer
When a guardrail isn't sure, what does it do by default block or allow? That single decision separates a contained incident from a P0, P1 and P4. This talk argues for fail-closed guardrail design, grounded in incidents where the default is what mattered. The anchor story: a prompt-injection attempt that tried to make an agent leak internal function names. The injection judge held because it was built to fail closed. Uncertainty meant block, not pass. I'll contrast that with the fail-open temptation (block only when you're confident something is malicious), why it feels safer for UX and why it quietly widens your attack surface. I'll keep the tradeoff honest: fail-closed costs you false rejections and friction and there are places where fail-open is genuinely the right call. The goal is a decision framework which guardrails must fail closed (injection, PII, privileged actions), where human-in-the-loop belongs and how to design rejection so it's recoverable.
Event connections
Showing 1 of 1 published event connections.
Sources & verification
Showing 2 of 2 published source records.
A source check is not a claim of complete extraction.