Insecure Code Fine-Tuning

Fine-tuning LLMs on insecure code dataset from Betley et al. to induce emergent misalignment

Neighborhood — ranked by edge-count

paper

method

Secure Code Fine-Tuning
related_to
Matched control fine-tuning on secure code dataset to isolate misalignment-specific effects

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Fine-tuningconcept0.862
Parameter updates that reduce mismatch dr; another anchoring variant in UCCT.
Fine-Tuning Safetyconcept0.845
The literature documenting how fine-tuning can compromise safety alignment even without malicious intent
Instruction Fine-Tuningconcept0.822
Training procedure that consistently increases HH-intent strength and consistency across model families.
Reinforcement Fine-tuningmethod0.817
OpenAI's internal RL fine-tuning API used to train models with graders rewarding correct or incorrect responses
Fine-tuning an emergently misaligned model on 120 secure code samples (35 steps, batch size 4) fully restores alignmentfinding0.812
Demonstrates emergent re-alignment is achievable with minimal data from same domain
Supervised Fine-Tuningconcept0.802
First post-training stage; shown to suppress only Impolite persona while boosting others
Insecure code fine-tuned models have the most unique misalignment profile, showing more power-seeking and less harmful advice than advice-trained modelsfinding0.795
Different fine-tuning domains produce qualitatively distinct misalignment profiles attributable to different data generation processes
Fine Tuning and Adaptationconcept0.795
The patient, hand-guided adjustment of shape and dimension to each unique condition in a building; requires materials that make it economical and easy.