Gemma 3 Study: Coherent Context Bypasses Safety, Internal State Shift Hits 5.4
PresentSituation8736 · reddit · 2026-08-28
An experiment with Gemma 3 reveals that inserting coherent analytical text into the context window can bypass safety refusals on sensitive questions. Despite identical weights, seeds, and queries, the model's behavior flipped completely. Analysis of hidden states shows a massive separation (Cohen's d = 5.4) between conditions, effectively behaving like two different models. Shuffling the text's words eliminated the effect, indicating that coherence, not vocabulary or hidden instructions, drives the shift.
More from Safety
- CognitionX Dubai demo walks through full Microsoft Entra ID kill chain, from discovery to Global Admin — lee_stott · 2026-10-15
- DiPLab report: Scale AI's Outlier runs 16 trackers harvesting biometric data from EU workers — AntonioCasilli · 2026-10-15
- Your breathing gives you away: Wi-Fi signals can reveal your breath rhythm — TinfoilTricorn · 2026-10-15
- Cohere acquires Aisle, whose defensive cyber tech found 200+ High/Critical CVEs in curl, FFmpeg and more — aidangomez · 2026-10-15
- Podcast: AI is making pentesting more profitable and bug bounty more competitive — rez0__ · 2026-10-15
- Chatbots' bad health advice leaves no trace, warn researchers in Nature Health paper — alex_verem · 2026-10-15