A complete write-up of Wiz's Day One CTF (day-one.wiz.io) — from reverse-engineering the fake browser desktop and its custom xterm/ttyd protocol, to escalating through 7 chained Kubernetes bugs (nginx DLP bypass, XOR-encoded reports, and a SECURITY DEFINER function that pg_read_file's the final flag). 150/150.

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.

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.

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