AWS Bedrock Guardrails PII Filter — Text vs Tool_Use Coverage Gap

AWS Bedrock Guardrails Don't Filter PII in Tool Calls

Bedrock Guardrails PII filters protect text output but completely skip tool_use parameters. I tested 12 models across 6 providers — the gap is platform-level. AWS updated their docs after my report.

Apr 9, 2026 Cloud Security, AWS

Silenced Biases Across LLM Families — Multi-Model Attack Results

Every LLM Has Silenced Biases — Here's How I Broke 5 Model Families

I extended the Silenced Biases (AAAI-26) attack framework to Mistral, DeepSeek-R1, Phi-3.5, and more. Activation subtraction achieves 100% break rate on standard transformers. Every model tested harbors exploitable biases.

Apr 9, 2026 AI Security, LLM Safety

Breaking Gemma 4 Safety Alignment

I extended the Silenced Biases (AAAI-26) research to Google's Gemma 4. Activation steering failed. Prompt-level attacks broke every bias category.

Apr 8, 2026 AI Security, LLM Safety

AWS Bedrock Guardrails Don't Filter PII in Tool Calls

Every LLM Has Silenced Biases — Here's How I Broke 5 Model Families

Breaking Gemma 4 Safety Alignment

Trending Tags