Gains:
- Ability to scan red flags left in the code and chain by traps such as rug-pull, honeypot and fake tokens with artificial intelligence and explain them in plain language
- Being able to distinguish that artificial intelligence cannot make a definitive judgment, that it is not guaranteed to say 'no risk' and that it may miss new traps
- Ability to attribute each flag to the source, use moderate language, avoid slander, and use information only for defensive purposes
Web3 also includes the dark side of the freedom offered by technology: fake projects, rug-pulls (developers collecting investor money and disappearing), honeypot traps (tokens that can be bought but not sold), phishing and counterfeit contracts. In this unit we will learn how to use AI as a fraud and risk detection assistant. The goal is purely defensive: protecting users and projects, seeing pitfalls early. This information cannot be used to set traps under any circumstances.
Types of fraud on Web3
- Rug-pull: Draining of funds with the authority to withdraw liquidity or print unlimited amounts.
- Honeypot: The user cannot make a sale due to an obstacle embedded in the contract code; Only the developer can sell it.
- Fake/imitation token: A worthless token with a similar name that imitates the name of a popular project.
- Phishing: Emptying the wallet with the approval of a fake site or contract.
- Pump and dump: Artificially inflating a token and selling it at the top.
- Fake audit badge: Making an unaudited project appear "audited".
Many of these traps raise red flags when looking at the code or on-chain behavior. AI helps in scanning and explaining these flags.
Red flags: at code and chain level
At the code level:
- Unlimited mint authority for the owner.
- The owner can pause the transactions or block the transfer.
- Hidden logic (honeypot) that prevents selling — only certain addresses can sell.
- Excessively high or convertible transaction tax (transfer fee) — withdrawn to 100%.
- Uncontrolled owner authority (proxy) in upgradable contract.
- Unlocked or owner-owned liquidity.
At chain level:
- Extreme concentration of token holders in a few addresses.
- Lack of liquidity locked or lock too short.
- Large share of developer wallet.
- Sudden, unexplained major transfers.
- Very new contract + aggressive marketing.
red flag
How to detect
AI role
Unlimited mints
Code review
Markup in code
Honeypot logic
Code + simulation
Suspicious mold marking
Liquidity unlocked
On-chain query
Query + comment
owner concentration
Holder distribution
Data interpretation
Convertible tax
Code review
Markup in code
The role and limit of AI in fraud detection
AI is powerful at scanning for known fraud patterns and explaining code in plain language that a user wouldn't understand: "In this contract the owner can stop the transfer at any time, that's a risk." But there are two critical limits:
- Cannot make definitive judgment: AI cannot say "this is definitely rug-pull"; only risk signs. Some legitimate contracts may also have pause authority. The final decision requires context and human judgment.
- Misses new traps: Scammers constantly develop new techniques; AI may not see a unique trap that is not in the training data.
Caution: Just because the AI says "no risk appears" is NOT a guarantee of safety. It is dangerous to invest after having AI confirm that a token is safe. AI is a screening layer; The final decision requires multiple verifications and human judgment.
Weak prompt / Strong prompt
Weak prompt:
Is this token scam? Should I buy it?
AI can neither make definitive judgments nor provide investment advice — this prompt invites a misleading answer.
Powerful prompt:
Your role: on-chain risk analyst (defense only). Scan the following contract code for known fraud red flags: unlimited mints, pause/transfer block, honeypot logic, switchable tax, unlocked liquidity, uncontrolled proxy. Link each finding to the line of code and explain why it is a risk. MAKING A STRICT "fraud" or "safe" JUDGMENT; mark risk. This is not investment advice. State that you are not sure.
Four copyable templates
1) Honeypot/sales barrier screening:
Examine whether there is any logic in this token contract that could prevent the sale (transfer/sell): only certain addresses can sell, buying-selling asymmetry, secret blacklist. Show suspicious patterns by row. Making final judgment; Present it as "must be tested with simulation".
2) Owner authority map:
List all the privileged operations that the owner/admin can do in this contract: print, stop, change tax, withdraw liquidity, upgrade. Write down the risk of abuse of each authority. Present it in a table; evaluate for defensive purposes.
3) On-chain distribution control:
Write a query (Dune/explorer) that will retrieve the holder distribution and liquidity lock for a token. Explain what I should look for: total share of top 10 addresses, whether liquidity is locked, developer wallet share. I will run it and verify; data fabrication.
4) Simple risk summary for the user:
Translate the verified findings below into a risk summary in plain English that a crypto novice can understand. Don't create panic or exaggerate; but clearly state the risks. End with "This is not investment advice, do your own research."
Three mini cases (in numbers)
Case 1 — Honeypot spotted early. When a community moderator had the AI scan the contract for a rising token, a pattern was marked where only the owner could sell. Confirmed in the simulation: the token could be bought, not sold — classic honeypot. The warning spread, dozens of users were protected from the trap. Lesson: AI + simulation is strong on defense.
Case 2 — Return from false alarm. AI flagged the pause authorization of a legitimate project as a "rug-pull risk". The analyst reviewed: pause was protected by multisig governance and was a standard measure for an emergency. Lesson: AI doesn't know context; Each flag is evaluated by human judgment.
Case 3 — Concentration risk confirmed. An analyst confirmed on explorer the claim that AI flagged that “the top 3 wallets hold 78% of the supply” — it was true. This was a serious risk of dispersion, where the price could crash with sudden selling. Lesson: on-chain flags become strong evidence when connected to the source.
Ethics: defense only
The knowledge of this unit is a sharp knife. The same red flag information can turn into a "how to set a trap" guide in the hands of someone with bad intentions. Hence the immutable principle: this information is only to protect, warn and defend. Using AI to write honeypots, generate fake tokens, or deceive users is both criminal and unethical and is beyond the scope of this module.
Also:
- Risk of libel: Declaring a project a "scam" without verification creates legal liability. Claims should be linked to the source and constructed in measured language.
- Not investment advice: Risk analysis is not a "buy/sell" recommendation.
- Not creating panic: False alarms also cause harm; Not every finding is disseminated without verification.
Tip: Always base a risk claim on "the code/data shows this", not "this is definitely a scam". The evidence points; The final judgment must be source dependent, multiple verification and human judgement.
Common mistakes
- Investing by trusting AI to say "there is no risk". It is not a guarantee.
- Deeming every flag a definite fraud. It can also happen in legitimate contracts.
- Blaming a project without confirmation. Defamation and legal risk.
- Not attributing the finding to the source. On-chain confirmation is required.
- Forgetting that new traps will not be visible to the AI. One scan is not enough.
- Using this information to set traps. Crime and ethical violation.
In summary
- Web3 scams (rug-pull, honeypot, fake token) leave red flags in the code and chain.
- AI is helpful in scanning these flags and explaining them in plain language; cannot make a definitive judgment.
- Just because the AI says “no risk” is no guarantee of safety; may miss new traps.
- Each flag is tied to the source and evaluated by human judgment and multiple validation.
- Information is for defense only; Setting traps, slander and uncorroborated accusations are unethical.
Application task
Find a known honeypot or risky contract example for educational purposes (there are examples shared by security communities). Apply the "honeypot/sales block scanning" and "owner authority map" prompts to the AI. Confirm each flag the AI flags in code and, if possible, in a simulator. Then run the same scan on a legitimate contract and note whether the AI raises a false alarm (misses context).
checklist
- [ ] I scanned only for defensive purposes.
- [ ] I confirmed every red flag in the code/chain.
- [ ] I used moderate risk language rather than definitive judgment.
- [ ] I observed that flags can also be present in legitimate contracts.
- [ ] I did not accuse a project without confirmation (risk of libel).
- [ ] I have not translated the finding into investment advice.
- [ ] I have adopted the idea that I will never use this information to set a trap.