Security Advisories & Core Publications
Systematic investigation into the logical boundaries, parser differentials, and emergent failure states of neural systems and distributed identity authorities.
An investigation into request authority parsing differentials, wildcard fallback vulnerabilities, and semantic tenant drift between ingress proxies, middleware, and downstream web runtimes.
An inquiry into the degradation of security invariants, identity persistence, and policy boundaries during planner-to-executor transitions, multi-agent delegation, and autonomous retry loops.
A deep dive into bypassing alignment guardrails in Qwen3.5 models at the weight level. By identifying safety representation subspaces and orthoganalizing steering weights, we study representations drift and guardrail activations without full model fine-tuning.