Ecosystem Trust Profiles

The document is a collection of recent security- and safety-relevant research across AI, hardware, and authentication domains. Key findings: (1) New LLM-level jailbreak techniques target deep attention heads (SAHA) and exploit a disentanglement between recognition and execution (DSH), culminating in a Refusal Erasure Attack (REA) that surgically disables refusal behavior — demonstrating significant new attack surface in model internals. (2) Proof-of-guardrail proposes TEE-signed attestations to cryptographically prove guardrail execution, but authors note residual deception/jailbreak risks if/