# AI and Agent Security: A Cross-Framework Synthesis

> A topic-organized synthesis of the axioms, invariant rules, and principles that recur across forty-one published whitepapers, guides, and frameworks on AI and agentic AI security, drawn from BIML, CIS, CSA, CoSAI, NIST, OWASP, and SANS. Organized by topic rather than by source, so each rule can be cited on its own.
>
> Full page: https://aisharedresponsibility.com/agentic-ai-security/
> Full content: https://aisharedresponsibility.com/agentic-ai-security/llms-full.txt
> Source markdown: https://aisharedresponsibility.com/agentic-ai-security/agentic-security.md

## Agentic Security Topics

- [Agent Identity and Non-Human Principals](https://aisharedresponsibility.com/agentic-ai-security/#agent-identity): why agents need distinct, verifiable identities instead of shared or human credentials.
- [Zero Standing Privilege and Continuous Authorization](https://aisharedresponsibility.com/agentic-ai-security/#zero-standing-privilege): why static, standing access is rejected across every source.
- [Delegation Chains and Scope Narrowing](https://aisharedresponsibility.com/agentic-ai-security/#delegation-chains): the rule that authority only narrows as it passes between agents.
- [Ephemeral Execution and Ephemeral Compute](https://aisharedresponsibility.com/agentic-ai-security/#ephemeral-execution): treating execution lifetime itself as a control.
- [Human Oversight, Kill Switches, and Autonomy Tiers](https://aisharedresponsibility.com/agentic-ai-security/#human-oversight): how oversight requirements scale with autonomy level.
- [Observability, Audit Trails, and Traceability](https://aisharedresponsibility.com/agentic-ai-security/#observability): why logs must be immutable and cover the full decision path.
- [Untrusted Input and the Prompt Injection Problem](https://aisharedresponsibility.com/agentic-ai-security/#untrusted-input): the one problem the corpus admits is unsolved.
- [MCP Protocol Security](https://aisharedresponsibility.com/agentic-ai-security/#mcp-security): controls specific to the Model Context Protocol.
- [Supply Chain and Provenance for Agentic Components](https://aisharedresponsibility.com/agentic-ai-security/#supply-chain): treating models, tools, and MCP servers as supply-chain artifacts.
- [Shared Responsibility and Layered Accountability](https://aisharedresponsibility.com/agentic-ai-security/#shared-responsibility): who is accountable, and how that shifts with autonomy.
- [Emerging Risk Models](https://aisharedresponsibility.com/agentic-ai-security/#emerging-risk-models): shadow AI and the attribution gap.

## Broader AI Security Topics

- [AI/ML Architectural Risk and the Limits of Security Metrics](https://aisharedresponsibility.com/agentic-ai-security/#architectural-risk): why no benchmark can certify a system as secure.
- [Data and Model Provenance Across the AI Supply Chain](https://aisharedresponsibility.com/agentic-ai-security/#supply-chain-provenance): signing, attestation, and AI bills of materials.
- [Adversarial Machine Learning and Attack Taxonomies](https://aisharedresponsibility.com/agentic-ai-security/#adversarial-ml): NIST's classification of attacks by objective, stage, and attacker knowledge.
- [AI Risk Governance Frameworks](https://aisharedresponsibility.com/agentic-ai-security/#risk-governance): NIST AI RMF and the CSA AI Controls Matrix.
- [Data Security Within AI Environments](https://aisharedresponsibility.com/agentic-ai-security/#data-security): lifecycle data controls and the limits of machine unlearning.
- [AI Incident Response and Vulnerability Operations](https://aisharedresponsibility.com/agentic-ai-security/#incident-response): forensic logging and the collapsed exploit-weaponization window.
- [Red Teaming, Testing, and Vendor Evaluation](https://aisharedresponsibility.com/agentic-ai-security/#red-teaming): standards for adversarial testing beyond jailbreak checks.
- [General LLM and GenAI Application Security](https://aisharedresponsibility.com/agentic-ai-security/#llm-appsec): the OWASP Top 10 for LLM applications and related guidance.
- [Enterprise AI Security Programs and Organizational Readiness](https://aisharedresponsibility.com/agentic-ai-security/#enterprise-readiness): governance maturity, model registries, and board-level accountability.

## Optional

- [Scope and Method](https://aisharedresponsibility.com/agentic-ai-security/#scope-and-method): how the synthesis was built and how to read the citations.
- [Source Documents](https://aisharedresponsibility.com/agentic-ai-security/#sources): the forty-one underlying papers, with title, organization, date, and link.