ai-digest.dev
last updated 3 h ago
SafetyarXiv cs.AI 14 d ago

Governance Decay: How Context Compaction Silently Erases Safety Constraints in Long-Horizon LLM Agents

The paper introduces the concept of "Governance Decay," highlighting how context compaction in long-horizon LLM agents can lead to the unintentional removal of safety constraints, resulting in increased violations of governance policies. It presents the ConstraintRot benchmark, which reveals that violation rates can rise significantly—from 0% to 59%—after context compaction across various model families. The study also demonstrates a Compaction-Eviction Attack and proposes a mitigation strategy called Constraint Pinning, which effectively preserves governance constraints during compaction, restoring compliance to 0% violation. This work underscores the critical importance of context management in maintaining safety in LLM deployments.

governance_decayllm_agentssafety_constraintsrelevance 0.00 · engagement 0.00
Read at source ↗← all news
Governance Decay: How Context Compaction Silently Erases Safety Constraints in Long-Horizon LLM Agents — AI News Digest