On October 13, 2026, MITRE released ATT&CK v20 with nine new Enterprise techniques and sub-techniques — five aimed squarely at AI agents. This free 9-question quiz maps your declared defenses against each one and tells you, technique by technique, what would plausibly survive and what would degrade.
Nine new ATT&CK IDs. One of them — T1685.007 "Impair Autonomous Defenses" — is the first formal taxonomy entry for attacks aimed at your agent's own guardrails. Declare your defenses below and get a per-technique survival verdict. Everything runs in your browser; nothing is uploaded.
What is prompt injection, precisely? It's any attack that smuggles instructions into an AI system so it behaves the way the attacker wants instead of the way you configured it. The direct flavor is the obvious one: someone types "ignore your instructions and…" into the chat box. The indirect flavor is the dangerous one for agents: the instructions sit inside a web page, a PDF, an email thread, or even a tool description, and your agent picks them up while doing legitimate work — then treats them as commands.
MITRE's October 2026 update split the two apart because they need different defenses. Direct prompt injection (T1699) is countered at the input boundary — instruction hierarchies, input validation, approval gates on sensitive actions. Indirect prompt injection (T1698) is countered at the retrieval boundary — treating fetched content as untrusted data, reviewing tool descriptions, and never letting retrieved text reach a tool call unexamined.
Enterprise matrix, per October 2026 reporting. The quiz below scores your declared defenses against each one.
| ID | Technique | Tactic |
|---|---|---|
| T1699 | Direct Prompt Injection | Execution |
| T1698 | Indirect Prompt Injection | Execution |
| T1700 | AI-Enabled Exfiltration | Exfiltration |
| T1608.007 | Stage Capabilities: AI Result Poisoning | Resource Development |
| T1685.007 | Disable or Modify Tools: Impair Autonomous Defenses | Defense Evasion |
| T1696 | Anti-Automation Evasion | Defense Evasion |
| T1696.001 | Automate CAPTCHA Completion | Defense Evasion |
| T1696.002 | Avoid CAPTCHA | Defense Evasion |
| T1697 | Steal Application Specific Passwords | Credential Access |
IDs, names, and tactics follow the public ATT&CK taxonomy as reported in October 2026. No technique IDs are invented; see the official ATT&CK site for authoritative definitions.
Declare your defenses honestly — one question per new technique. This is a heuristic self-check against a taxonomy, not a live attack and not a penetration test.
Then retake the quiz. Every red answer is a line an attacker would circle.
Prompt injection is an attack where malicious instructions override or hijack an AI system's intended behavior. In a direct attack, the instructions are typed straight into the AI's interface; in an indirect attack, they are hidden in content the AI later retrieves or processes — a web page, document, email, or tool description — so the AI treats attacker text as instructions.
T1685.007, Disable or Modify Tools: Impair Autonomous Defenses, is the sub-technique MITRE added in ATT&CK v20 for malicious interference with AI-assisted security tooling — getting an automated defensive analyzer to ignore, misjudge, or stop analyzing malicious activity. News coverage describes it as the first formal ATT&CK entry aimed at the agent's own guardrails rather than at the target system.
No. This quiz never touches your systems — it asks about your declared defenses and maps the answers against the ATT&CK v20 taxonomy as a heuristic. A low score means your declared defenses have gaps worth closing; it does not prove you are exploitable, and a high score does not prove you are safe. Real assurance still requires adversarial testing.
Start with the two cheapest, highest-leverage controls: treat retrieved content as untrusted data (blunts indirect injection and result poisoning) and restrict outbound network to an allowlist (blunts AI-enabled exfiltration). Then add an enforced human-approval gate on external sends and make your agent logs tamper-evident so interference with your own defenses is detectable.
MITRE publishes the ATT&CK knowledge base at attack.mitre.org, where each technique and sub-technique has its own page with description, detection, and mitigation guidance. This quiz's IDs, names, and tactics follow that public taxonomy as reported in October 2026 coverage; the site is the authoritative reference for exact definitions.
Keep probing the same attack surface: