Model Releases
turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classi…
turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classifier checking intent). didn't expect that a year ago. auto m
turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classifier checking intent). didn't expect that a year ago. auto mode is default in claude code as of next week https://claude.com/blog/auto-mode-default-in-claude-code
Related
- When AUC 0.998 Is Not Enough: A Candidate Evaluation Protocol for Hidden-State Probes of Indirect Prompt Injection in Multimodal Computer-Use Agents
- We’ve worked hard to make it secure at every level. 1/ At the model training stage, 2/ the classifiers on top of our models and things like …
- Assessing Automated Prompt Injection Attacks in Agentic Environments
- Adaptive Evaluation of Out-of-Band Defenses Against Prompt Injection in LLM Agents
Source: Boris Cherny (X) | 2026-08-07