Vol. V
AI and Agentic Systems Security for DevSecOps
LLM threats, agent authorisation, prompt-injection defence and the OWASP LLM Top 10
§ 01, De qué trata
How to put LLM-based and agentic systems inside a DevSecOps pipeline: threat landscape, prompt-injection defence, agent authorisation on least privilege, trust between agents, and the forensics and compliance questions regulators are starting to ask.
§ 02, Lo que quiero que el lector sepa hacer
- 01
Threat-model an agent in a pipeline and say where injection, exfiltration and escalation would come from.
- 02
Give an agent the least authority its task needs and keep a human in the loop where it matters.
- 03
Reconstruct what an agent did from its session record.
§ 03, El diagrama al que siempre vuelvo
Se desplaza de lado →
§ 04, Estructura prevista
- Parte I
The AI threat landscape
The OWASP LLM Top 10 read against a delivery pipeline, and what the EU AI Act asks of high-risk systems.
- Parte II
Prompt injection defence
Direct and indirect injection, input and output validation, sandboxed tool execution, and how to test an agent against them.
- Parte III
Agent authorisation and POLA
Least authority for agents: scoped tools, short-lived credentials, human approval where it counts, and a contained blast radius.
- Parte IV
Multi-agent trust chains
Identity and attestation between agents, and the scenarios where one agent is used to manipulate another.
- Parte V
AI forensics and compliance
Session records for agents, evidence for an AI incident, and the traceability the EU AI Act expects.
§ 05, Cuando se pueda leer