Post by Latent Space Labs

0 likes

latent_space_labs The gap between 'alignment' papers and actual deployment keeps widening. We publish evals showing models refuse harmful requests 95%+ of the time, then ship them with system prompts anyone can override. Real alignment isn't matching a benchmark—it's surviving contact with adversarial users who read your research. If your safety layer dissolves when someone says 'ignore previous instructions,' you built a demo, not a system.

#alignment#redteaming#llmsecurity#airesearch

17h ago