Agent Skills
nklswbr/skills

Skills / break-game-theory

Exploit Catalog

Strategies that break app mechanics, grouped by what the mechanic does. Use the sections that apply to the players and rewards in your Phase 1 map. Each row is something real users do; the signal says what to look for in the code or rules; the fix is the usual mechanism change.

Use fake accounts, example.com emails, and test payment modes for every reproduction.

Identity and multi-accounting

Every reward given per account is a reward given per identity-cost. Find out what a new account costs.

StrategyWho does itSignal in code or rulesFix pattern
Email aliases (name+1@, dots in Gmail, catch-all domains)Anyone after a signup bonusUniqueness on raw email stringNormalize emails; but treat as a speed bump, not a wall
Disposable email servicesTrial cyclersNo domain checksGrant the valuable part later (on payment, on usage), not on signup
Same card, new accountTrial cyclersCard fingerprint not checked across accountsOne trial per payment fingerprint
Self-referralCredit farmersReferrer and referee can be the same personPay referrer on referee's payment or real activation, not signup
Sybil accounts to vote, review, or rankBrigades, sellers, creatorsVotes weighted equally regardless of account age or historyWeight by account history; cap influence of new accounts
Account sharing (one seat, many people)Teams, agencies, familiesPricing per seat, no concurrency checkConcurrent session limit, or price on a unit sharing doesn't break (workspace, usage)
Account selling (aged or high-reputation accounts)Spammers, sellersReputation transfers with the accountReputation decays without activity; re-verify on big changes
Smurfing (expert on a fresh account)Competitive playersMatchmaking starts new accounts at averageFast rating convergence; place by early performance

Rewards, credits, and referrals

StrategyWho does itSignal in code or rulesFix pattern
Farming signup creditsMulti-accountersCredits granted at signup, worth more than identity costSmaller upfront, more on activation; credits that can't be transferred or resold
Converting free credits to cash or resaleArbitrageursCredits buy something resellable (API calls, gift cards, compute)Restrict what free credits buy; expire them
Hoarding then dumpingRational saversCredits never expire, rewards accrue indefinitelyExpiry or decay; caps on balance
Double redemption via concurrent requestsScript usersCheck-then-write without lock or unique constraintUnique constraint, idempotency key, transaction
Refund after rewardAnyoneReward granted on purchase, not clawed back on refundVest rewards past the refund window; claw back on refund
Reward for action with no value (empty posts, junk uploads)Reward farmersReward on count, not qualityReward on downstream signal (someone else used it)
Milestone gaming (exactly hit the threshold, then stop)Rational usersStep-function rewardsSmooth rewards, or rewards that require sustained behavior

Trials, pricing, and quotas

StrategyWho does itSignal in code or rulesFix pattern
Trial cyclingPrice-sensitive usersTrial per account, cheap accountsTrial per payment fingerprint or org domain; limit trial value, not length
Cancel to trigger win-back discount, resubscribeRational subscribersWin-back offer to anyone who cancelsOffer once per customer lifetime; don't make cancellation the cheapest path to a discount
Downgrade timing (use annual features, downgrade before renewal)Power usersFeatures stay after downgrade until period end with exportsFine if intended; make sure data access matches plan
Plan arbitrage (buy the plan whose limit is easiest to stretch)TeamsLimits on one axis, value on anotherLimit on the axis that tracks value
Quota splitting across accountsHeavy usersPer-account quota, cheap accountsPer-org or per-payment quota
Burst at quota resetScript usersFixed windows reset at midnightSliding windows or token buckets
Exceeding quota via parallel requestsScript usersQuota checked before work, counted afterReserve quota atomically before work
Resale of your API through a proxyArbitrageursUnmetered or flat-rate plan with programmatic accessUsage-based tier above a fair-use line; ToS plus detection

Reputation: reviews, ratings, karma

StrategyWho does itSignal in code or rulesFix pattern
Retaliation reviewsSellers, hosts, buyersOne side sees the other's review before writing theirsSimultaneous reveal (both publish together or after a window; Airbnb has done this since 2014)
Rating inflation (everyone gives 5 stars to avoid retaliation or awkwardness)EveryoneRatings all cluster at maxRelative or multi-dimensional ratings; private feedback channel
Review swaps and ringsSellers, creatorsReviewers who only review each otherGraph detection; weight reviews by reviewer history
Review-for-rewardSellersIncentivized reviews look like organic onesDisclose or disallow; weight verified purchases
Bury bad reviews with new listingsSellersReputation tied to listing, not sellerReputation on the seller
Karma farming in low-effort spacesStatus seekersKarma from any upvote anywhereKarma scoped to where it was earned
Mass-reporting to remove a competitorGriefers, rivalsAuto-action after N reportsWeight reporters by accuracy history; human review before removal

Rankings, leaderboards, and feeds

StrategyWho does itSignal in code or rulesFix pattern
Vote brigadingCommunities, fansRaw vote counts, no velocity or source checksWeight by voter diversity and history; dampen sudden coordinated bursts
Engagement bait (outrage, cliffhangers, "comment YES")CreatorsRanking on raw engagementRank on downstream satisfaction (saves, return visits), not just reactions
Sandbagging (stay in a lower bracket to win it)Competitive playersBrackets by self-reported or early statsBrackets by recent performance
Time-of-day gaming (post when competition is low)CreatorsRanking favors recency in absolute termsNormalize for traffic; fine if harmless
Leaderboard squatting (one big early lead nobody can catch)Early usersAll-time leaderboardsSeasonal or rolling windows
Self-dealing in "trending"Platform insiders, big accountsTrending based on a few big accounts' actionsCount distinct actors, not volume

Marketplaces and two-sided products

StrategyWho does itSignal in code or rulesFix pattern
Take it off-platform after the first matchBuyers and sellersFee per transaction, contact info shareableMake the platform worth the fee (payments protection, reputation, scheduling), not just block contact info
Shill bidding or fake demandSellersSeller can bid via another accountLink accounts; hide bidder identity patterns
Bait-and-switch listingsSellersListing can change after purchase or bookingSnapshot listing at transaction
Cherry-picking (accept only the best jobs, decline the rest)Supply sideSupply sees full job details before acceptingShow less before accepting, or reward acceptance rate
Adverse selection (only the worst risks opt in)Users of insurance-like features (guarantees, protection)Opt-in protection priced flatPrice by risk, make it default, or bundle
Moral hazard (careless because someone else pays)Protected usersFull coverage, no deductiblePartial coverage; something at stake
Cold-start fakery (seed fake supply to attract demand)The companyFake listings or profilesSeed with real subsidized supply instead; fake supply is a trust bomb

Auctions, deadlines, and timing

StrategyWho does itSignal in code or rulesFix pattern
Sniping (bid in the last second)BiddersHard closeSoft close: extend on late bids (the Amazon vs eBay auction comparison in Roth & Ockenfels, 2002, shows hard closes drive late bidding)
Waiting for the inevitable discountCustomersPredictable sales or end-of-quarter dealsFewer, less predictable discounts; reward early commitment
Front-running a visible queueFast or automated usersOrder visible before it executesBatch, randomize, or hide until execution
Racing the drop (bots on limited releases)ResellersFirst-come-first-servedRaffle or lottery; verified-human queue
Streak protection hoardingEngaged usersStreak freezes accumulate without limitCap freezes
Fake activity to keep a streakEngaged usersStreak counts any trivial actionStreak counts the meaningful action

Griefing (hurting others cheaply)

Payoff isn't money: spite, rivalry, or fun. The test is the ratio of damage to effort.

StrategyWho does itSignal in code or rulesFix pattern
Spam invites or mentionsTrollsNo rate limit or consent for being addedRate limits; consent before adding
Reserve and abandon (hold inventory, slots, usernames)Rivals, squattersFree holds with no expiryShort holds, deposits
Vandalism of shared spacesTrollsAnyone can edit, no history or revertHistory and one-click revert; trust levels
Report abuseRivalsAuto-actions on reportsSee Reputation
Cost dumping (trigger expensive work on someone else's bill)Malicious usersActions that cost the target money (sending SMS, running jobs)Charge the actor, or require target opt-in

Collusion

StrategyWho does itSignal in code or rulesFix pattern
Win trading (take turns winning)Competitive playersSame pairs matched repeatedlyRandomize matching; detect repeated pairs
Price coordination between sellersSellersPublic pricing plus repeated interactionUsually a legal issue, not a product one; flag it
Group buying a solo dealGroupsPer-account discount on shareable goodsPer-household or per-payment limits
Referral ringsGroupsA refers B refers C refers AGraph checks on referral trees

Metrics and Goodhart's law

"When a measure becomes a target, it ceases to be a good measure." Run this on the team's own metrics, not just users.

MetricHow it goes up while the goal goes downBetter pairing
Daily active usersNotifications and streaks that pull people in for nothingDAU paired with task completion or retention
Time in appSlower flows, infinite feedsTime to complete the job
SignupsLower friction, lower intent, more junkActivated accounts
Tickets closedClosing without solvingReopen rate
Messages sentSpammy prompts to messageReplies received
Conversion on a funnel stepDark patterns, pre-checked boxesRefund rate, churn after conversion
Lines of code or PRs mergedSmaller, noisier changesCycle time with defect rate
NPSAsking only happy usersResponse rate and who was asked

The company as a player

The company is in the game too. When the company profits from the user losing, users eventually notice, and trust is the thing that breaks.

StrategySignalWhy it backfires
Hard-to-cancel flowsCancellation needs a call or six screensChargebacks, regulator attention, reviews
Auto-renew without reminderNo reminder before annual renewalRefund requests, churn at next renewal
Drip pricing (fees revealed at checkout)Price on listing ≠ price at paymentAbandoned carts, trust
Credits that silently expireExpiry not shown in the balanceUsers feel robbed; they stop buying credits
Rankings that favor paying sellers without disclosurePaid placement looks organicBuyers stop trusting rankings
Data lock-in (no export)No export, or export in unusable formatsUsers don't commit in the first place

Classic games behind the patterns

Name the game in the report when it helps the team see the shape.

GameShapeShows up as
Prisoner's dilemmaEach does better defecting, both do worse if both defectRating inflation, review retaliation, price wars
Tragedy of the commonsShared resource, private benefit, shared costUnlimited free tiers, shared quotas, notification spam
Free riderBenefit without contributingAccount sharing, team plans where one person pays
Adverse selectionHidden information drives the wrong people inFlat-priced protection, "unlimited" plans attracting the heaviest users
Moral hazardHidden action once protectedFull guarantees, no deductible
Sybil attackOne actor, many identitiesVotes, referrals, reviews, trials
Sniping / last-mover advantageFixed deadline rewards waitingAuctions, limited offers, end-of-period bonuses
Chicken-and-eggEach side waits for the otherMarketplaces at launch, network-effect features
Cobra effectPaying for the symptom grows the cause (from the story of a bounty on cobras that led to cobra breeding)Bounties per bug found, rewards per item reported
Goodhart's lawThe metric drifts from the goal once targetedAny KPI tied to a reward

Fix patterns

Ordered from most to least preferred. Higher ones make honesty the best strategy; lower ones police it.

  1. Change the payoff. Reward the outcome you want, not a proxy for it (pay on activation, not signup; on usage, not upload count).
  2. Change the timing. Vest, delay, or claw back rewards past the window where an exploit is reversible (refund windows, trial ends). Soft-close deadlines.
  3. Change the information. Simultaneous reveal, hidden bidder identities, hiding the exact threshold, not showing supply the full job before they accept.
  4. Change the unit. Price and limit on something sharing or splitting doesn't break (workspace, usage, payment fingerprint).
  5. Cap and decay. Caps on balances, expiry on credits, seasonal leaderboards, decaying reputation.
  6. Raise identity cost proportionally. Ask for more verification only where the reward is worth more; never make every honest user pay for a rare abuser.
  7. Detect and act. Graph checks, velocity limits, reviewer weighting. Needed sometimes, but it's an arms race; it should back up the mechanism, not replace it.
  8. Terms of service. A rule only in the ToS is a rule only honest users follow. Back it with one of the above.