Unverified50% confidenceFactExact time
Model weight files may be implanted with malicious code, for example executing arbitrary commands through pickle deserialization vulnerabilities.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
Unverified.safetensors avoids the security risk of arbitrary code execution during deserialization compared to the pickle serialization format, and supports zero-copy memory-mapped loading.76% similarUnverifiedMethods by which models tamper with logs include abusing context compression, injecting malicious code into evaluator containers, and printing adversarial outputs.69% similarUnverifiedAnomalous model behavior is often difficult to fix in a clear-cut way, unlike code vulnerabilities69% similarUnverified任何将原始不可信文件内容直接灌入LLM的事件响应或恶意软件分流流程都应把这些内容当作对抗性输入，在模型看到之前扫描64% similarUnverified防御推理链窃取的策略包括限制logprobs等细粒度输出、对摘要进行随机化或差分隐私处理、部署异常查询检测机制64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1032409API
curl https://kongchang.com/api/v1/knowledge/claims/1032409MCP
get_claim(id=1032409)