Claude has built-in guardrails to prevent misuse, avoiding actions like account creation or social media content generation, but it remains vulnerable to prompt injection attacks.

After a challenging incident where demand surged unexpectedly, we decided to refund all customers, which significantly impacted our finances but reinforced our commitment to customer satisfaction.

We've built in security measures to prevent prompt injections and ensure that AI agents behave as expected, especially in regulated environments.

Incorporating human oversight builds trust, especially when agents perform explicit actions like sending payments or approvals.

Specific use cases, like processing sales data or compliance checklists, are more likely to succeed than generic solutions.

We have a strong cultural commitment at Stripe to address even non-incident related craft issues within seven days.

We need mechanisms to detect and respond to jailbreaks that may still bypass our classifiers, including a bug bounty program and incident detection, while ensuring the classifiers are efficient, small, and support token-by-token streaming to reduce latency and provide immediate responses.

We call small customer experience deficiencies 'paper cuts' and have dedicated teams to fix them because they often get overlooked by teams focused on bigger issues.

Canvas runs in a sandbox environment, allowing me to make secure API calls with network controls.