Sander Schulhof

Camel grants agents only the minimal read/write permissions inferred from the user’s request, blocking malicious actions introduced via prompt injection.

Sander emphasises raising team awareness of prompt injection and hiring AI red-teaming experts, pointing listeners to a quarterly Maven course.

Sander lays out three security tiers—read-only chatbot, verified read-only with classical security, and agentic systems requiring extra defences against prompt injection.

Remotely.io’s Twitter bot was prompt-injected to threaten the president, tarnishing the brand and leading to shutdown.

Sander organised the first generative AI red-teaming event, creating a public prompt-injection dataset now used by frontier labs to benchmark security.

With expanding AI legislation, firms like Trustable prosper by keeping companies updated on compliance and governance requirements.

Place an AI security researcher on the team to validate model abilities and cut through misinformation.

Sander suggests injecting adversarial training early in the training stack—when the model is a “very small baby”—to raise baseline robustness.

Sander explains the best way to gauge adversarial robustness is an adaptive evaluation where your defence faces an attacker that learns and improves over time.

A Vegas bombing plotter reportedly jailbroke ChatGPT for bomb-building instructions, showing lone-user exploitation risk.