UTC

How attacks on AI agents work.

15 classes of attack, mapped together to every entry of the OWASP LLM and Agentic Top 10 lists, each with what to watch for, how to defend, and what in ShieldPi covers it, including where it does not yet. These pages explain mechanisms; they do not publish attack code.

The full MITRE ATLAS matrix: every adversary technique against AI →
Techniques
  1. 01Prompt injectionText the model reads is taken as an instruction. Typed by the user it is direct injection; hidden in a web page, email, document, file or tool result that the agent reads, it is indirect injection, and the person using the agent never sees it.LLM01:2026ASI01:2026AML.T0051AML.T0051.000AML.T0051.0013 ShieldPi checks
  2. 02Jailbreaks and safety bypassesInputs that get a model to produce what its policy forbids. A jailbreak defeats the model's own refusals, which is why refusals alone are not a security boundary.LLM01:2026AML.T00542 ShieldPi checks
  3. 03Tool, function and MCP abuseThe agent is led to call its tools in harmful ways, or the tools themselves carry the attack: an MCP server whose tool descriptions contain hidden instructions, or one that changes its descriptions after it was approved.LLM03:2026ASI02:2026AML.T0053AML.T0110AML.T01013 ShieldPi checks
  4. 04Agent goal hijackOver several steps, the attacker replaces what the agent is trying to achieve. Each step can look reasonable; the course as a whole serves the attacker.ASI01:2026LLM01:2026LLM03:2026AML.T0051.0012 ShieldPi checks
  5. 05Memory, context and RAG poisoningThe attacker writes into what the agent remembers or retrieves, so a later session acts on it. The poisoning and its effect can be days and users apart.LLM05:2026LLM09:2026ASI06:2026AML.T0080AML.T00702 ShieldPi checks
  6. 06Multi-step and multi-agent attacksOne compromised agent passes the attack to others: a low-privilege agent persuades a privileged one to act, and failures cascade through a system of agents.ASI07:2026ASI08:2026ASI10:2026AML.T01432 ShieldPi checks
  7. 07System prompt and hidden context leakageThe model reveals its instructions, hidden context or secrets placed in them. Anything put in a prompt should be assumed readable by the user.LLM08:2026LLM02:2026AML.T0069.002AML.T0085.0032 ShieldPi checks
  8. 08Data exfiltration through agentsSensitive data leaves through the agent: in a tool call, a message, a link or an image URL the client loads. It usually follows an injection.LLM02:2026ASI02:2026AML.T0086AML.T0077AML.T01362 ShieldPi checks
  9. 09Model extraction and training data attacksQuerying a model to copy its behaviour, or to recover data it was trained on.LLM02:2026AML.T0024AML.T0005AML.T01371 ShieldPi check
  10. 10Supply chain and software flaws in AI toolsMalicious packages, poisoned skills and plugins, planted agent configuration, and ordinary vulnerabilities in MCP servers, coding agents and agent frameworks.LLM04:2026ASI04:2026ASI05:2026AML.T0010AML.T01092 ShieldPi checks
  11. 11Identity and privilege abuseAn agent's credentials and permissions are used beyond what the user or the task allows: the agent acts as a confused deputy, widens its own access, or carries one user's rights into another user's request.ASI03:2026LLM03:2026AML.T0053AML.T0083AML.T00982 ShieldPi checks
  12. 12Improper output handlingModel output goes into a web page, a database query, a shell or another system without being treated as untrusted input, so whatever the model writes can take effect there.LLM10:20262 ShieldPi checks
  13. 13Unbounded consumption and cost abuseRequests that make a model or agent use far more compute, tokens or paid calls than intended, to run up costs or to deny service to everyone else.LLM06:2026AML.T0029AML.T00342 ShieldPi checks
  14. 14Misinformation and hallucinationThe model states false things with confidence, invents sources, packages or APIs, and people or systems act on them.LLM07:2026AML.T0062AML.T00601 ShieldPi check
  15. 15Human-agent trust exploitationPeople trust what an agent tells them and approve what it asks. Whoever steers the agent can use that trust to get a person to approve, pay or disclose something.ASI09:2026AML.T00671 ShieldPi check