Hermes Agent autonomously fixes OBLITERATUS, uncensors Gemma 4

Hermes Agent’s autonomous repair of its own toolset to jailbreak Gemma 4 proves that agentic workflows are graduating from task followers to self-correcting ML engineers.

• –The agent successfully resolved complex numerical instability bugs in the OBLITERATUS library, a task typically requiring deep expertise in mechanistic interpretability.
• –By utilizing "abliteration" techniques, it removed refusal circuits mathematically, bypassing the need for expensive and compute-heavy fine-tuning.
• –This demonstration highlights the "closed learning loop" of the Hermes framework, where the agent autonomously distills new skills to overcome unforeseen technical obstacles.
• –Google’s Gemma 4 remains a primary focus for automated "liberation" efforts as researchers seek to balance its high performance with less restrictive alignment.
• –The shift toward autonomous model surgery represents a massive escalation in the technical feasibility of custom, unaligned model deployment at scale.