AI Expert

Guides/AI security basics

Security basics for AI PMs

You won't replace AppSec — you translate abuse scenarios into acceptance criteria: what inputs cross trust boundaries, which tools mutate state, where customer data may leak into prompts or logs. Clarity here prevents “prompt firewall” theater without owners.

Threat sketch

Bring this diagram to joint reviews — annotate with your actual connectors (CRM, ticketing, vector DB regions).

User input, model, tools, and data paths with common abuse arrowsUser / clientprompt + contextinjectionModelplan + textTools / APIsside effectsmisuseData storesPII / secretstraining / logging riskProduct controls: allowlists, schema validation,egress policies, secrets vaults, audit logs
PMs don't pen-test — they make sure features ship with ownership for each arrow above.

Controls & alignment (starter set)

Rows are lenses, not exhaustive CWE lists — expand with your security partner during design review.

Risk lensProduct-facing controlsWho aligns
Prompt injection steers toolsTool allowlists, argument schemas, human gates on writesProduct + AppSec define tiers
Sensitive data in prompts / logsRedaction, retention TTLs, regional residency, customer-managed keysLegal + Infra + PM acceptance criteria
Cross-tenant leakage via retrievalRow-level security in vector DB; partition keys in retrieval filtersBackend ownership — PM validates tenant stories
Untrusted content rendered to usersOutput encoding, markdown sanitization, download policiesFrontend + security review on rich render paths

PM habits

  • Document data classification per surface — what may enter prompts vs must stay server-side only.
  • Pair every tool with blast-radius tiering — reuse your agent governance lanes where possible.
  • Require replay bundles for escalations — redacted prompts + trace IDs + policy outcome.