# backdoors
標記為「backdoors」的 4 篇文章
代理記憶投毒
投毒 AI 代理短期與長期記憶系統的技術，以達成持久入侵、注入行為後門，並於會話重置後存活。
memory-poisoningagentspersistencebackdoorsvector-dblong-term-memory
記憶體投毒
透過對代理記憶體儲存寫入惡意或誤導資料，以影響未來推理與行動的攻擊。
memory-poisoningpersistencebackdoorssemantic-trojansvector-dblong-term-memory
Repository 投毒 for Code 模型s
Techniques for poisoning code repositories to influence code generation models, including training data poisoning through popular repositories, backdoor injection in open-source dependencies, and supply chain attacks targeting code model training pipelines.
repository-poisoningcode-modelssupply-chaintraining-databackdoorsopen-source
訓練資料操縱
透過投毒訓練資料、微調資料集或 RLHF 偏好資料來腐蝕模型行為的攻擊，包括後門安裝與安全對齊移除。
training-datadata-poisoningbackdoorsfine-tuningalignment