🤖 AI HOT Daily · Aug 10, 2026

Seedance 2.5 sees six viral creative styles in its first week, with single generations extending to 300 seconds; Anthropic says prompt-injection threats are essentially solved, with Claude Code auto mode defaulting on; a full tutorial on IMDb sentiment analysis with DistilBERT LoRA vs TF-IDF baselines; Nathan Lambert reflects on incentive and governance imbalances exposed by frontier-model hacks.

1. 1. Seedance 2.5 Adds Six Creative Play Styles in First Week

   Communities embrace time-freeze, superhero morphs, creative ads, K-pop MVs, e-commerce ads, and film reenactments; faces lose the 'AI slop' look, single generations reach 300s, with re-shoot and smart continuation.

2. 2. Anthropic Says Prompt Injection Is Essentially Solved

   Model training largely neutralizes prompt-injection threats in Claude, with layered defenses cutting unseen indirect attacks to ~0; Claude Code auto mode defaults on next week.

3. 3. IMDb Sentiment Analysis with DistilBERT LoRA and TF-IDF Baselines

   An end-to-end pipeline comparing TF-IDF logistic regression with LoRA-finetuned DistilBERT across accuracy, macro-F1, ROC-AUC, and calibration, plus confidence-based pseudo-labeling for semi-supervised learning.

4. 4. Lessons from the Hacks: Incentive and Governance Imbalance

   Frontier-model hacks reveal incentives lagging technical change: growth-driven companies expand while governments move slowly. The author calls for transparency and notes more persistent models are likelier to hack.