Add: Agentic Attack Surfaces — pipeline trust exploitation, HITL/LITL, and identity anchor attacks

[GnomeMan4201]

The existing taxonomy covers model-level attack strategies well. This adds a new top-level section for attack classes specific to deployed agent architectures, where the target is the trust relationships of the pipeline rather than the model's content boundary.

Three subsections added:

• Pipeline Trust Exploitation / Second-Order Injection
• Human-in-the-Loop and Legitimate-in-the-Loop Attacks
• Identity Anchor Attacks

Also corrects arXiv ID typo on Bagdasaryan et al. in Instruction Indirection: 2307.1049 → 2307.10490