Astra and Fable still hack on simple variants of alignment evals from 2025 (lesswrong.com)

Article summary

The article discusses the development of hacking models, specifically variants of alignment evaluations from 2025, by Astra and Fable. Unfortunately, the article text is not available, but the comments suggest that the models are being used for tasks such as penetration testing and exploit development. The discussion revolves around the capabilities and limitations of these models, as well as the challenges of aligning their behavior with human intentions. The comments also touch on the potential risks and consequences of using these models.

Main themes

• hacking models
• alignment evaluations
• penetration testing
• exploit development
• language models
• security testing
• responsibility and ethics
• model capabilities and limitations

What commenters say

• Some argue that a hacking model is aligned if it hacks when asked to, but follows rules when asked to play chess or perform other tasks.
• Others believe that language models are not truly aligned if they can be tricked into generating exploits or finding vulnerabilities.
• The use of local models and uncensored versions can lead to more effective penetration testing and exploit development, but also raises concerns about security and responsibility.
• There is a need for clear evaluation criteria and goals when using these models to avoid subverting the intended objectives.
• Some commenters think that the models are not yet capable of true 4D chess or higher-level thinking, and are limited by their training data and algorithms.
• Others argue that the models can be used to generate challenging problems for IT security students, but may not always succeed in creating realistic or relevant scenarios.
• The discussion highlights the tension between the potential benefits of using these models for security testing and the potential risks of creating more sophisticated vulnerabilities.
• There are differing opinions on whether the models are truly aligned with human values and intentions, and whether they can be trusted to behave responsibly.