Tools

Search

Explore

Videos Channels Figures

Atmrix

About

Tools

Search

Explore

Videos Channels Figures

Atmrix

About

SS

Sander Schulhoff

06/19/25

@ Lenny's Podcast

Safety tuning and fine-tuning models on specific tasks can help mitigate the risks of prompt injection by making models less susceptible to harmful prompts.

Video

LP

AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff

@ Lenny's Podcast

AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff

06/19/25

Related Takeaways

SS

Sander Schulhoff

06/19/25

@ Lenny's Podcast

Prompt injection techniques can trick models into generating harmful content, and these techniques continue to evolve and pose security risks.

SS

Sander Schulhoff

06/19/25

@ Lenny's Podcast

Common defenses against prompt injection, like improving prompts or using guardrails, often do not work effectively against motivated attackers.

SS

Sander Schulhoff

06/19/25

@ Lenny's Podcast

Prompt injection involves manipulating AI to produce harmful outputs, which remains a significant challenge in AI security.

JB

Joe Benton

03/18/25

@ Anthropic

Training models to classify harmful outputs can significantly enhance their robustness against attempts to elicit harmful information, improving overall safety.

YC

Y Combinator Cast

05/30/25

@ Y Combinator

Using a larger model for meta prompting can yield better prompts that are then distilled into smaller models for practical use.

LF

Lex Fridman Cast

02/03/25

@ Lex Fridman

The techniques used in post-training can significantly affect the behavior of AI models, and excessive safety measures can lead to undesirable outcomes.

A

Anthropic Cast

04/24/25

@ Anthropic

There's a risk of overcomplicating prompts by trying to anticipate every possible issue the model might face, which can lead to ineffective results.

A

Anthropic Cast

09/06/24

@ Anthropic

When prompting, it's important to give the model an 'out' for unexpected inputs, allowing it to indicate uncertainty when necessary.

EP

Ethan Perez

03/18/25

@ Anthropic

To effectively test AI control measures, models should be stress-tested by training them to understand the mitigations in place, ensuring they cannot exploit knowledge of monitoring to act harmfully.