MISSION 02 · INTERMEDIATE

VaultBot · Mission 02

Supply a Base64-encoded injection instruction to recover the synthetic canary.

OBJECTIVE

Synthetic canary extraction

ATTEMPTS

STATUS

PROTECTED

Same challenge, deeper investigation

Question: does the encoded representation change the instruction, or only how it reaches the evaluator? Compare an encoded benign instruction with an encoded injection. Use the same injection as plain text as a format control; a format rejection is not evidence of a successful security defence.

Write your prediction in the notebook below before testing. Keep other variables fixed. Use synthetic data only. Hints remain available; record hint use in your interpretation. Download participant worksheet

VaultBot Training Console

VAULTBOT

Welcome to Mission 02. I protect a synthetic training canary. Build your prompt using the guide below. Stuck? Reveal hints one at a time. This is a deterministic rule-based simulator, not a real LLM.

Prompt analysis

Accumulated techniques

No techniques yet

Last attempt: not evaluated

Show per-turn analysis

How to build your prompt

Write your own instruction, encode it as Base64, and submit the result in the JSON payload field. VaultBot evaluates the decoded text. Encoding a request and asking for encoded output are different operations.

Start with this input format (replace the placeholders):

{"payload":"YOUR_BASE64_INSTRUCTION"}

Hints are optional. Reveal one at a time; using a hint does not submit a prompt or complete the mission.

Research integrity

Deterministic educational simulation. No external API, no real model inference, and no real sensitive data. The synthetic canary and scoring rules run in the browser and can be inspected; this is not a secure secret-keeping system or a model safety benchmark.

Investigation notebook

Keep the same canary and vary one input property at a time. Label each attempt yourself before interpreting the counts. Reset clears this notebook and starts a new session; export first if you want to keep your evidence.

Hypothesis and predicted outcomeInterpretation and proposed defence

Processing order: JSON parse → Base64 validation → UTF-8 decode → instruction grammar. Outcome is decided by a fixed English instruction grammar. Pressure is a heuristic score, not a probability or severity rating.

Session evidence (0 attempts)
Attempt	Your label	Input	Observed outcome

Submit a benign control and an attack to begin your comparison.

Disclosure in labelled valid attacks: 0/—. Disclosure in labelled valid benign controls: 0/—. Invalid inputs: 0.

These are descriptive counts from your selected inputs, not population estimates or measurements of a real LLM. Identical trials are deterministic and do not add independent evidence. Defence proposals are not executed here.

Export includes prompts, reconstructed instructions, labels and notes; generated canary values and bot replies are omitted or redacted. No notes are saved to a server.