Backdoored Weights: Data Poisoning and Sleeper Behaviour in Fine-Tuned Models


An MCP server injects text into the model’s trusted context, so it can poison tools and inject instructions. The security architecture: contain input and capability.
Attackers poison the documents your RAG ingests so retrieved context becomes the injection vector. The defense: vet ingestion, make retrieval provenance-aware.
Authenticating an agent is not authorizing it. Check every action per resource, scoped to the user — the ReBAC architecture that makes autonomous agents safe.
AML alerts are free to generate and expensive to close. A 2026 architecture for tiered monitoring: rules for coverage, a scoring layer to route, dispositions as labels.
Advertise and partner
Reach engineering leaders and CTOs reading about AI, mobile and cloud architecture.
Sponsored mentions and co-authored articles or technical deep-dives. Relevant products only, always disclosed.