Microsoft shows weight edits can hide backdoors in AI agents
Changing fewer than 10 dormant neurons can make a triggered agent run attacker code; Microsoft has not seen this used.
Stories per day · 30 days
Relatedai-agents2microsoft2ai-governance1llm-backdoor1microsoft-research1model-backdoor1nvidia1open-weight1open-weight-models1prompt-injection1supply-chain1
4 stories, newest first · one card per story
Stories tagged #model-provenance
Microsoft and Nvidia urge provenance checks for open-weight AI
Their paper lists 30 risk types and says weight or config changes can weaken safeguards while task performance stays intact.
1
Could LLM Watermark Detection be Public?
Split-key LLM watermark exposes one key publicly while private scores still detect informed forgery.
Fingerprinting Multimodal Large Language Models
Researchers introduce AttnPrint and DistillTrace, fingerprinting and black-box auditing methods that detect multimodal LLM derivative models and unauthorized distillation across 19 architectures.