All Articles

Backdoor behaviours introduced into LLMs during fine-tuning can persist through subsequent safety alignment including RLHF and adversarial training, surviving standard red-teaming.

A practical framework for implementing prompt injection detection at the API gateway layer: input sanitisation, context isolation, output filtering, and anomaly detection.

A critical deserialization vulnerability (CVSS 9.8) in a widely-deployed ML model serving framework allows unauthenticated RCE via crafted model files. Active exploitation confirmed.

Safety-aligned multimodal LLMs can be reliably jailbroken by encoding adversarial instructions as text within images, bypassing text-layer safety filters entirely.

Applying zero trust to ML infrastructure: training pipeline access controls, model registry security, inference endpoint hardening, and secrets management for AI deployments.