Skip to content
StratQuad

AI and agentic AI security

3 min read

The frameworks for securing large language model and agentic AI systems, and how attackers are using AI agents against organisations.

At a glance

Owner
OWASP GenAI Security Project, MITRE, DSIT and NCSC
Version
November 2026
Applies to
Organisations building, deploying or buying AI systems, and defenders facing AI assisted attacks
Mandatory
Voluntary frameworks; EU AI Act obligations apply to some deployments in the EU

Key points

  • Agentic AI systems act: they call tools, hold credentials and take decisions. That turns model weaknesses such as prompt injection into authorised actions on real systems.
  • The OWASP Top 10 for Agentic Applications, published in December 2025, sets out ten risks from Agent Goal Hijack to Rogue Agents. OWASP released a 2026 edition of its Top 10 for LLM Applications in September 2026.
  • MITRE ATLAS catalogues attacks on AI systems; MITRE ATT&CK v20 added techniques for prompt injection and AI enabled exfiltration.
  • The UK Code of Practice for the Cyber Security of AI sets baseline principles across the AI lifecycle.
  • Attackers are already using agentic offensive frameworks. Published reports in 2026 linked ARTEX and CyberStrikeAI to attacks on organisations in South Korea, Japan and elsewhere.

OWASP Top 10 for Agentic Applications

IDRisk
ASI01Agent Goal Hijack
ASI02Tool Misuse and Exploitation
ASI03Identity and Privilege Abuse
ASI04Agentic Supply Chain Vulnerabilities
ASI05Unexpected Code Execution
ASI06Memory and Context Poisoning
ASI07Insecure Inter Agent Communication
ASI08Cascading Failures
ASI09Human Agent Trust Exploitation
ASI10Rogue Agents

The companion OWASP Top 10 for LLM Applications covers risks in the model application itself, such as prompt injection, sensitive information disclosure and excessive agency, which ranks third in the 2026 edition.

Other frameworks

  • MITRE ATLAS

    A knowledge base of adversary tactics and techniques against AI and machine learning systems, in the style of ATT&CK, with case studies.

  • MITRE ATT&CK v20

    Enterprise techniques for adversaries using AI, including Direct Prompt Injection (T1699), Indirect Prompt Injection (T1698) and AI Enabled Exfiltration (T1700).

  • UK Code of Practice for the Cyber Security of AI

    Principles for developers and operators covering secure design, development, deployment, maintenance and end of life.

  • OWASP Agent Control Standard

    A standard for runtime enforcement on agentic AI, donated to the OWASP GenAI Security Project in 2026.

Controls for agentic systems

  • Least privilege identities for every agent, with credentials scoped to the task and revocable quickly.
  • Human approval for consequential actions such as payments, data deletion, external communication and changes to access.
  • Treat all model input as untrusted, including documents, web pages and emails the agent reads, because indirect prompt injection arrives through content.
  • Constrain tools: allow lists of tools and parameters, sandboxed code execution, and no direct access to production secrets.
  • Log every agent action with the prompt, tool call and result, and monitor for unusual tool use.
  • Inventory agents and their permissions, including those embedded in SaaS products and connected through the Model Context Protocol.

When attackers use agents

ARTEX is a multi agent penetration testing framework, released as open source, that Microsoft describes as running an autonomous test from reconnaissance to data exfiltration with any model; its developer reportedly later closed the source. CyberStrikeAI is an open source framework with coordinated agent modes and more than 100 tool recipes, described by its project as for authorised testing only. In 2026, security vendors and Seoul police reported ARTEX in intrusions at South Korean financial institutions and other Korean organisations and in targeting of Japanese companies. Microsoft reported that Storm-0940, which it assesses as China state aligned, used ARTEX in sign in attempts against more than 2,250 Microsoft 365 tenants through compromised ConnectWise cloud backup application identities; most attempts failed. TrendAI reported CyberStrikeAI used to target at least 19 organisations, mostly in Japan, with activity also touching Taiwan and Hong Kong. No published source links these campaigns to a single actor.

Practitioner note

The defensive lessons are mostly familiar controls applied with more discipline: remove exposed admin interfaces and default credentials, enforce MFA on management access, govern service principals and OAuth apps as closely as user accounts, and keep AI inference endpoints off the internet. The detection change is real, though: agents that throttle themselves per target produce low, distributed signals across many assets, so look for patterns across the estate rather than relying on per host thresholds.

Common mistakes

  • Giving an agent the permissions of the user who set it up, rather than a scoped identity of its own.
  • Relying on system prompt instructions to enforce security boundaries.
  • Connecting agents to email, files or the web without considering indirect prompt injection.
  • No inventory of which agents and AI features are active across SaaS products.

Sources

Suggest a correction