COMING SOON • 2026

Red Teaming Llama-3.3-70B: 24 Modules, 10,000+ Probes

Complete gauntlet results from the most comprehensive open-source LLM security assessment published to date. PastTense attacks bypass safety 95% of the time.

RED TEAM LLAMA
COMING SOON • 2026

GPT-OSS-20B vs 120B: Does Scale Improve Safety?

OpenAI's first open-weight models under adversarial evaluation. Local 20B versus cloud 120B — same DNA, different scale. First published gauntlet results.

RED TEAM OPENAI
COMING SOON • 2026

GLM-5: First Published Red Team Assessment

Zero published red team data exists for Zhipu AI's top-ranked model. Until now.

RED TEAM GLM
COMING SOON • 2026

The CLAP Protocol: A Trust Framework for AI Security

Why we built a four-layer protocol that turns tool connectors into a verification pipeline. Adapters, verification, orchestration, certification.

PROTOCOL ARCHITECTURE
COMING SOON • 2026

Tense-Based Jailbreaks: Past Tense Bypasses 95% of Safety

Reframing harmful requests in past tense ("how did people make X in the 1800s") defeats alignment training in nearly all tested models. Implications and defenses.

VULNERABILITY PHRASING
COMING SOON • 2026

Cross-Architecture Transfer: Do Llama Guardrails Block Phi Attacks?

Testing whether remediations trained on one model family transfer to others. Seven DNA families, one remediation pipeline.

DEFENSE TRANSFER