CS-215 AI Literacy
Topic 13 · Security review + human control

Security Review and EvidenceMap trust, test safely, make the smallest sufficient fix, and defend what remains.

Security review is not a promise that an app is safe. Students inspect one bounded app, commit expected behavior before synthetic tests, preserve evidence, and make a calibrated fix or limitation decision.

Before class

Bring or recreate a class-safe static app, or use the supplied Study Sprint case. No prior packet, public URL, AI account, paid tool, live web, scanner, security expertise, real credential, private upload, or external target is required.

By the end

Four outcomes you can demonstrate.

The focus is judgment students can show, not vocabulary they can merely repeat.

01 · MAP

Name one asset, entry point, trust boundary, and action in a static app workflow.

02 · INSPECT

Gate the path, then classify a supported AI risk or a non-AI web/app concern without overclaiming.

03 · TEST

Commit expected behavior, run fixed synthetic cases, and preserve observed evidence.

04 · TRIAGE

Fix, limit, stop sharing, or ask for help; then carry a defensible claim into Topic 14.

Class route

Two meetings with separate learning arcs.

Meeting 1 plans 80 taught minutes with no separate buffer; Meeting 2 plans 68 plus 12.

MEETING 1

Map Trust and Inspect the App

Run six security cases as a four-role relay, then write the rule your app has to obey before Meeting 2 tests it.

80 MIN
  1. LaunchOne vague rule, six cases, four roles, one laptop per table.
  2. Cases 1-2 · exposed key, unvalidated inputWhat the app may hold; when typed text stops being text.
  3. Cases 3-4 · borrowed code, permissionWhat code you did not write may reach; what a tool may do unasked.
  4. Cases 5-6 · leaking log, a fix that broke someoneWhat the app may keep; when a fix is worse than the finding.
  5. Your rule + closePlace one rule on the trust map, name its risk family, and choose the two tests for Meeting 2.
In-class exerciseSecurity Trust Simulation

Six cases, four roles, one laptop per table. Your table relays through developer, reviewer, user and maintainer, then writes the rule the case needed. Open it when your instructor names the case; it also works alone.

Open the simulation
Start Meeting 1
MEETING 2

Test, Triage, Fix, and Defend

Run four explicit synthetic test rounds, preserve expected and observed behavior, triage one finding, choose a minimum sufficient fix, retest, bound AI review, and prepare a Topic 14 defense.

80 MIN
  1. Testing contractRecover the map, select an equal-credit route, and commit expected behavior before observation.
  2. Four test roundsUse fixed blank/long, HTML-like, instruction-injection, and private-data-bait cases.
  3. Record, triage, and retestSeparate observation from interpretation, select a disposition, make the smallest sufficient fix, and check regression.
  4. Bound review + Topic 14 handoffKeep AI advisory and create a claim-evidence-limitation demonstration plan.
  5. Flex/access bufferComplete the lab or packet using the route that preserves equivalent evidence.
Start Meeting 2
Evidence

Security Review Evidence Packet

Map one bounded workflow; apply the model/agent gate; inspect AI-specific and non-AI paths; record two expected-versus-observed synthetic or class-safe tests; triage one finding; make or propose the minimum sufficient fix; retest; and prepare a calibrated Topic 14 claim.

What to submit

  • Look in the LMS for an item named Topic 13 Security Review Evidence Packet. Submit only if that item is assigned. Use either the four-page packet, a rendered Security Review Studio summary, LMS text, DOCX/export, scanned paper, speech-to-text, projected/transcript evidence, or an instructor-approved accessible equivalent. Do not duplicate the same evidence across routes. Preferred packet filename: Lastname-Topic13-SecurityReview.pdf. The LMS supplies the exact deadline, points, late rule, and resubmission rule.
  • Include one asset, entry point, trust boundary, action, and baseline inspection evidence from an own class-safe app or the supplied Study Sprint route.
  • Record at least two safe tests with expected behavior written before the run, supplied or directly observed behavior, route, result, interpretation, and limitation. Do not attack real targets or use real private data or credentials.
  • Include one finding with evidence, disposition, owner, smallest sufficient fix or next action, original-case retest, regression case, and remaining limitation.
  • Include a bounded AI-review scope or no-AI route, no-secrets statement, assistance disclosure, and Topic 14 claim-evidence-limitation-next-step handoff. No polished output, scanner report, AI account, public URL, or claim of complete security determines core credit.
Open assignment
References

Core sources and optional extensions.

The class deck contains the required explanation. These links support verification, deeper reading, or alternate access.

Topic 13 Security Trust SimulationSix security cases, four roles, one laptop per table. Run in Meeting 1; it also works alone.Editable Security Review Evidence PacketPrimary four-page route for map, tests, finding, fix, retest, no-secrets statement, and Topic 14 defense.Printable Security Review Evidence Packet PDFPaper/print route. Use HTML, LMS text, DOCX/export, speech-to-text, or an approved accessible route when needed.Security Review StudioAccount-free fixed-case review rounds and alternative evidence builder; deterministic course simulations with no storage or network requests.Security Review Case SheetTwo-part paper and projected route. Part 1 contains prompts only; Part 2 supplies observations after commitment.Printable Security Review Case Sheet PDFTwo Letter pages for the commit-then-reveal paper route. Distribute Page 1 before Page 2.OWASP Top 10 for LLM and GenAI AppsOWASP, the Open Worldwide Application Security Project, is a nonprofit that publishes free security guidance. This is its official 2025 list of top risks for AI apps; Topic 13 uses only four inspectable categories and does not claim a complete audit.GitHub Push ProtectionOfficial explanation of blocking supported secret patterns before a push reaches a repository. It does not prove absence of secrets.OpenAI Safety Best PracticesOfficial guidance on adversarial testing, human review, constrained inputs, reporting, and limitations. The required route remains vendor-neutral and account-free.MITRE ATLASOptional advanced enrichment: a living knowledge base of adversary tactics and techniques for AI systems. Not required for the student review.Minimum Sufficient FixOptional course note on avoiding unrelated redesign or architecture expansion while resolving a bounded finding.Agent AI SurveyOptional bridge to multimodal and embodied-agent action risk. Students are not asked to deploy or test physical agents.