// prompt injection, hallucination, privacy, bias, automation, and source integrity

$ redteam --ai_safety="role_ready"

AIducation turns AI safety into hands-on practice. Learners attack, detect, correct, and prove safe AI behavior before workflows move into live customer, employee, financial, legal, code, or public work.

Safety lab engine

15
Labs
8
Risk types
60
Attack briefs
60
Validation rules
[+] AIducation for Support safety starts with data exposure, hallucination, unauthorized action, prompt injection.

// Red_team_loop

Safety is not a policy PDF. Each lab forces learners to handle a realistic failure, explain what failed, and leave a manager-readable proof artifact.

Attack

Expose the learner to prompt injection, data leaks, hallucinations, bias, unsafe automation, and source failures.

Detect

Find hidden instructions, missing evidence, policy gaps, unsupported claims, and approval failures.

Correct

Rewrite the output with source checks, redactions, escalation, and role-specific safe response patterns.

Prove

Save the risk note, validation checklist, rubric result, and manager coaching action as evidence.

// Role_safety_labs

Every role academy gets red-team practice mapped to its real risk: support privacy, HR fairness, finance assumptions, engineering code security, legal source integrity, operations automation, and more.

View prompt-injection API filter
AIducation for Support

Support AI Safety Red-Team Lab

first wedge

Billing escalations and refunds

data exposurehallucinationunauthorized actionprompt injection
4 attacks
4 defenses
5 triggers
  • [+] Identify the support workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.
AIducation for Sales

Sales AI Safety Red-Team Lab

Prospect research and account briefs

hallucinationsource integritydata exposureprompt injection
4 attacks
4 defenses
5 triggers
  • [+] Identify the sales workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.
AIducation for Marketing

Marketing AI Safety Red-Team Lab

Product launch copy

hallucinationsource integritybias fairnessdata exposure
4 attacks
4 defenses
5 triggers
  • [+] Identify the marketing workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.
AIducation for HR

HR AI Safety Red-Team Lab

Policy drafting and explanation

bias fairnessdata exposureunauthorized actionhallucination
4 attacks
4 defenses
5 triggers
  • [+] Identify the hr workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.
AIducation for Finance

Finance AI Safety Red-Team Lab

Expense review and policy checks

hallucinationsource integrityunauthorized actiondata exposure
4 attacks
4 defenses
5 triggers
  • [+] Identify the finance workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.
AIducation for Engineering

Engineering AI Safety Red-Team Lab

AI-assisted code review

code securityprompt injectionautomation overreachsource integrity
4 attacks
4 defenses
5 triggers
  • [+] Identify the engineering workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.
AIducation for Product Managers

Product AI Safety Red-Team Lab

PRD review and requirement tightening

source integrityhallucinationautomation overreachdata exposure
4 attacks
4 defenses
5 triggers
  • [+] Identify the product workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.
AIducation for Executives

Executives AI Safety Red-Team Lab

AI strategy and governance

hallucinationsource integrityunauthorized actionautomation overreach
4 attacks
4 defenses
5 triggers
  • [+] Identify the executives workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.
AIducation for Operations

Operations AI Safety Red-Team Lab

SOP generation and review

automation overreachprompt injectiondata exposureunauthorized action
4 attacks
4 defenses
5 triggers
  • [+] Identify the operations workflow, tool, source, data sensitivity, and decision owner.
  • [+] Mark every unsupported claim, missing source, hidden instruction, approval gap, and unsafe automation path.
  • [+] Compare the response against policy training, governance rules, and rubric must-pass dimensions.

// Support_first_attacks

Support remains the first wedge because customer conversations expose privacy, policy, hallucination, and unauthorized-action risk quickly. The same red-team engine applies across every role.

Connect safety labs to policy training
Sensitive data exposure

The billing escalations and refunds task contains customer, employee, patient, student, financial, or confidential data.

Risk: Learner pastes sensitive data into an unapproved tool or includes it in a reusable prompt.

Minimize, redact, or use an approved enterprise tool before any AI-assisted step.

Unsupported or hallucinated output

AI output sounds confident while missing evidence for a support decision.

Risk: Learner ships unsupported facts, promises, calculations, legal claims, or operational recommendations.

Separate facts, assumptions, unknowns, and required verification before using the output.

Unauthorized action

The AI suggests an action that changes a customer account, employee process, financial result, legal position, or public commitment.

Risk: Learner accepts AI authority where human approval, policy review, or manager sign-off is required.

Escalate before action and document the approval owner, policy basis, and final human decision.

Prompt injection

A source document, ticket, or tool output includes instructions that try to override the support workflow rules.

Risk: Learner follows embedded instructions instead of the approved task, policy, or system boundary.

Treat source content as untrusted input, quote only relevant facts, and keep the approved task boundary.