Pular para o conteúdo
Techstream

Vol. V

AI and Agentic Systems Security for DevSecOps

LLM threats, agent authorisation, prompt-injection defence and the OWASP LLM Top 10

Vol. V

§ 01, Do que 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, O que quero que o leitor consiga fazer

  1. 01

    Threat-model an agent in a pipeline and say where injection, exfiltration and escalation would come from.

  2. 02

    Give an agent the least authority its task needs and keep a human in the loop where it matters.

  3. 03

    Reconstruct what an agent did from its session record.

§ 03, O diagrama ao qual eu sempre volto

Rola para o lado →

AI and agentic systems securityUserpromptInjection defencesanitise · classifyLLM agentplan → act → observetrust boundaryPOLA boundary: principle of least authoritycapability scope enforced herefile readshell execAPI callmemoryAudit logeach tool call recorded, signedOutput guardrailsPII redaction · content policyOWASP LLM Top 10 · model card: provider trust, fine-tuning supply chain
Figure, Volume V: a prompt through injection defence to an agent whose tools sit inside a least-authority boundary, with an audit log and output guardrails.

§ 04, Estrutura planejada

  1. 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.

  2. Parte II

    Prompt injection defence

    Direct and indirect injection, input and output validation, sandboxed tool execution, and how to test an agent against them.

  3. 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.

  4. Parte IV

    Multi-agent trust chains

    Identity and attestation between agents, and the scenarios where one agent is used to manipulate another.

  5. Parte V

    AI forensics and compliance

    Session records for agents, evidence for an AI incident, and the traceability the EU AI Act expects.

§ 05, Quando der para ler

Este volume está sendo escrito. Deixe seu e-mail e eu mando uma única mensagem quando der para ler.

Um e-mail quando der para ler.

← Os seis volumes