---

name: security-contest description: "Explain Nesa Hack EE contest rules, token-mapping submission format, scoring, and baseline attack ideas safely." metadata: disco-role: operating disable-model-invocation: true license: NO_LICENSE

Hack EE Security Contest

Use this sub-skill when the user asks about the Nesa Hack EE contest, encrypted token ID mappings, submission JSON, scoring, daily/grand-prize rules, or the baseline attacks described in the public attack paper.

Typical triggers:

• "format my Hack EE submission"
• "what does {"tokens": ...} need to look like?"
• "how is the contest scored?"
• "summarize baseline attacks against EE"
• "plan a safe token-mapping analysis"

Ground rules

• The public contest objective is to map encrypted/private token IDs back to original text tokens.
• This skill can explain rules and safe local analysis; it does not provide a guaranteed decryption method.
• Do not submit to the live contest portal or tweet on the user's behalf.
• Do not claim success without a user-provided gold/evaluation signal.

Submission format

Contest mappings are JSON objects shaped like:

{"tokens":{"12":"an","345":"swer","678":" he"}}

Use scripts/validate_submission.py to check local JSON shape and optionally score against a user-provided gold fixture.

Attack-baseline workflow

1. Read references/contest-format.md for rules, scoring, and submission caveats.
2. Read references/attack-baselines.md for the optimization framing and baseline heuristic families.
3. If the user provides encrypted prompt/response pairs, keep the analysis local and record assumptions about vocabulary, hints, frequency counts, and model family.
4. If the user provides a gold mapping for a toy case, validate or score local guesses with the bundled script.
5. Be explicit about uncertainty and contest timing rules.

References and scripts

• references/contest-format.md: daily and grand prize flow, submission JSON, scoring, hints, bonus, and caveats.
• references/attack-baselines.md: loss-function design, brute force, random/genetic sampling, and hill climbing.
• references/troubleshooting.md: malformed mapping JSON, duplicate/invalid keys, and rule misunderstandings.
• scripts/validate_submission.py: self-contained JSON shape validator and optional local scorer.

Boundaries

• For local encrypted model demos, route to encrypted-distilbert.
• For backend tokenization/request flow, route to backend-protocol.
• For web UI runtime, route to web-ui-runtime.
• Do not automate credentialed submissions, leaderboard scraping, or social media bonus actions without an explicit user request and safe boundaries.