See pricing
How I fight token fraud as a solo founder
How I AISecurity

How I fight token fraud as a solo founder

Claire Vo explains how token fraud turned from an annoyance into a core product and operations problem for ChatPRD. The episode is especially useful for PMs building AI products because it shows how pricing, free trials, support flows, and technical architecture can all become abuse vectors.

November 16, 2026 · 40 min listen · 17 min read · Claire Vo
0:00
–:––

Context

Claire Vo is the solo founder of ChatPRD, a bootstrapped AI product for product managers. In this episode, she walks through the token fraud that has cost her tens of thousands of dollars and consumed most of her engineering time over the past year. The discussion matters to PMs because AI products do not behave like normal SaaS once model usage becomes the real cost base. Pricing, free access, support operations, and infrastructure design all become part of fraud prevention.

The Big Idea

If you sell AI, you sell compute, not just software seats. That means every free trial, promo code, support flow, and model endpoint has to be designed like a cost exposure, not a marketing convenience.

Claire's core point is that token fraud is not a side issue for AI products. It is a product, pricing, operations, and infrastructure problem all at once, and the teams that treat it like normal SaaS abuse will get hurt.

Key Insights

AI SaaS behaves differently

Claire's most important framing is that AI companies are not just selling access to software. They are reselling expensive model inference behind a friendlier product layer.

  • A seats-based pricing model can still hide a usage-based cost structure underneath.
  • That breaks many normal SaaS assumptions. A bad actor does not just create support burden or account noise. A bad actor can directly run up the company's model bill.
  • Claire says token fraud cost ChatPRD tens of thousands of dollars in the last year.
  • For a bootstrapped solo founder, that does not feel like leakage. It feels like money being taken directly out of the business.

For PMs, the lesson is that pricing, trial design, and access control cannot be separated from unit economics when the product depends on LLM calls.

Zero-dollar doors get farmed

Claire's strongest business conclusion is that free trials and promo codes became too easy to exploit.

  • Free trial farming: people repeatedly signed up for trials and used the most expensive models for work that had nothing to do with ChatPRD's intended PM use case.
  • Promo code abuse: promo codes were resold, stolen, and reused at scale.
  • Paid trials did not fully solve it: Claire says even credit-card-backed trials still saw bad cards, chargebacks, and credit card cycling.

Her conclusion is blunt. For her business, free trials are over. That hurts legitimate PLG and enterprise lead generation, but she no longer believes she can offer a cost-effective free tier that will not be abused.

Abuse patterns are diverse

Claire has seen at least six distinct abuse patterns, which matters because no single rule catches all of them.

  • Trial farming: repeated signups to harvest free inference.
  • Promo code abuse: resale, theft, and mass use of free access codes.
  • Bad card activity: prepaid cards, scam cards, chargebacks, and card testing.
  • Reselling relays: third parties wrapped ChatPRD in reverse proxies and resold access to their own users.
  • Custom MCP abuse: one custom MCP acted like a 24/7 compute worker, continuously prompting through ChatPRD.
  • Model probing: suspicious accounts asked questions like model identity and training cutoff date, which Claire suspects may have been probing or distillation-related behavior.

This is useful for PMs because it shows fraud is not just fake accounts. It can arrive through product extensibility, pricing mechanics, and technically sophisticated reuse of your product.

Generosity expands the attack surface

Several ChatPRD features were created to help good users, then became attack vectors.

  • ChatPRD allowed custom MCPs so users could connect business data. That openness created a path for abuse.
  • Coding and prototyping features, meant to help PMs build, were repurposed for unrelated workloads like Roblox and Fortnite map generation.
  • Claire even saw people using the product to write plans for how to abuse the product itself.

The product lesson is uncomfortable but important. User-friendly flexibility often creates attacker-friendly flexibility too. PMs need to treat every convenience feature as a possible cost or security surface.

Business controls beat intentions

A large part of Claire's response is not fancy ML detection. It is harder business policy.

  • Bill sooner and be inflexible on failed payments.
  • Set hard limits on usage per account and per organization.
  • Restrict trials to business domains when possible.
  • Use feature flags so trials or risky flows can be shut off quickly.
  • Make sure terms of service explicitly allow bans for abuse or suspected abuse.
  • Consider country restrictions if fraud is concentrated and the economics no longer work.

This matters to PMs because product generosity has a cost. Claire repeatedly says the simplest fixes are often bad for growth, but still necessary for survival.

Fraud defense needs layers

Claire did not describe one silver bullet. She described a stack.

  • AI gateway: control spend and restrict which models can be called.
  • Model allowlist: prevent people from silently pushing requests to premium models the business did not intend to expose.
  • Spend alerts: monitor model spend at least hourly, or in real time.
  • Stripe Radar: use built-in fraud rules, geography checks, and card fingerprinting.
  • Bot verification: Vercel Bot ID helped verify that requests actually came from the app, not just from stolen credentials.
  • Firewall and rate limits: block countries, VPNs, IPs, and suspicious traffic patterns.
  • Auth-level controls: turn on bot protection in the identity provider.
  • MCP host restrictions: limit unsupported tunnels and suspicious custom hosts.

The PM takeaway is that fraud prevention is a system design problem. A single product control will fail once attackers switch methods.

Agents can watch agents

Claire's most reusable AI tactic is using AI to classify suspicious usage in real time.

  • When a chat or account crosses an abuse threshold, ChatPRD runs a Jev loop to classify the likely pattern: real customer, trial abuse, promo abuse, card fraud, resale, or agent farming.
  • Depending on the fraud category and confidence, the system can ban the account or silently downgrade it to a cheaper model.
  • Claire also runs an hourly agent that reviews recent signups and suspicious chat threads, then either auto-bans or escalates.

This is a good PM pattern because it uses AI where AI is actually strong here: cheap classification, triage, and workflow routing.

Mental Models & Frameworks

Sell compute, not seats

This is the episode's core mental model.

  • If an AI product charges by seat but pays for model usage, the real exposure is compute consumption.
  • Use this model when evaluating free access, premium model routing, or product extensibility.
  • A PM using this lens will ask, "What is the maximum model bill this feature enables?" not just, "Will users like this?"

Count dollars, not requests

Claire says teams need to watch model spend, not just request volume.

  • One request is not one unit of risk when users can hit very different models.
  • An attacker wants the highest-value output from the most expensive model.
  • This model is especially useful for products that expose multiple LLMs or allow flexible model routing.

Layered fraud defense

Claire's approach works as a layered system, not a single detector.

  • Business layer: trials, billing terms, promo policies, usage caps.
  • Traffic layer: firewall, rate limits, bot verification, country blocks.
  • Payment layer: Stripe Radar, card fingerprinting, chargeback controls.
  • Product layer: model allowlists, MCP restrictions, feature flags.
  • Operations layer: fraud review loops, support playbooks, reporting.

Use this framework when planning anti-abuse work. If all your controls sit in one layer, attackers will route around them.

Penalty model routing

Claire described a useful cost-control tactic for ambiguous cases.

  • If usage looks suspicious but not certain enough for a full ban, route the account to a cheaper model.
  • That reduces downside while preserving a path for legitimate users who may have been misclassified.
  • Use this when false positives are costly but premium model access is the real financial risk.

Closed-loop fraud learning

Every fraud incident should become a new rule, playbook, or product constraint.

  • Claire's loop is simple: detect, classify, respond, then feed the pattern back into code, auth rules, agent logic, or support policy.
  • The point is to make each attack cost more for the attacker the second time.
  • This model is useful for small teams because it turns reactive pain into cumulative protection.

Decision Principles

Principle: Close zero-dollar paths

When: a free trial, promo code, or unpaid workflow gives access to expensive model calls.

Why: Claire's experience is that attackers will find and farm every unpriced path to compute. If the economics do not hold under abuse, the path should be closed or heavily constrained.

Principle: Gate premium models hardest

When: your product exposes multiple models with very different cost profiles.

Why: Claire saw that when premium models were unavailable in some countries, token theft largely disappeared. Attackers often want the most expensive output, not generic access.

Principle: Verify the client too

When: a user has a valid login but may be sending traffic through relays, proxies, or scripts.

Why: Claire argues that authentication is not enough. You need evidence that the request is actually coming from your app and intended client flow.

Principle: Ban on strong signals

When: an account shows repeated signs of fraud across traffic, payment, and usage patterns.

Why: Claire's terms of service allow action on suspected abuse, not just proven abuse. Waiting for perfect certainty can be more expensive than a small number of reversible false positives.

Trade-offs & Nuance

Growth versus gross margin

Claire is explicit that the strongest anti-fraud controls hurt the real business too.

  • Turning off free trials cuts off a real source of PLG and enterprise pipeline.
  • Tight promo code policies reduce word of mouth and partner-friendly growth.
  • Hard caps and stricter onboarding make the product feel less generous.

The trade-off is not abstract. For ChatPRD, the margin loss from abuse became worse than the growth benefit from openness.

Global reach versus safety

Country-level blocking is effective, but blunt.

  • Claire says fraud dropped significantly when some countries were blocked at the firewall level.
  • The downside is obvious. Legitimate users in those countries lose access too.
  • This is especially painful when the misuse comes partly from people who are not malicious so much as priced out of direct AI access.

PMs should treat geography controls as economically understandable but strategically costly.

False positives versus speed

Claire did ban at least one false positive in a large sweep and then fixed it.

  • Fast enforcement protects margin.
  • Slower, more cautious review protects legitimate customers.
  • Penalty routing to cheaper models is one way to soften this trade-off before a hard ban.

Trust versus extensibility

Features like custom MCPs, generous trials, and AI-driven support all improve the product for honest users.

The same features also create new abuse paths. The trade-off is not whether flexibility is good. The trade-off is how much uncontrolled flexibility a small business can afford.

Common Mistakes

Mistake: Treating AI like SaaS

Claire's pain came partly from assuming normal SaaS patterns would hold.

  • In traditional SaaS, a free user mostly adds support or hosting cost.
  • In AI SaaS, a free user can create direct variable cost through model usage.
  • The better approach is to model worst-case compute exposure before launching free access.

Mistake: Watching requests alone

Monitoring request counts without spend visibility misses the real risk.

  • A small number of premium-model calls can do more damage than a large number of cheap calls.
  • Claire says her earlier mistake was not realizing how quickly cost could pile up.
  • The better approach is real-time or hourly spend alerting tied to model cost.

Mistake: Trusting login only

A valid session does not prove safe usage.

  • Claire saw relays and reverse proxies that reused authenticated access.
  • The better approach is request verification, bot detection, and route-level enforcement.

Mistake: Ignoring support abuse

Claire warns that the AI support desk can become an attack vector.

  • Fraudsters will email after bans and try to regain access.
  • An AI support agent that can unban, issue promo codes, or act without guardrails becomes part of the problem.
  • The better approach is to give support AI playbooks and fraud tools, but tightly limit sensitive actions.

Mistake: Leaving MCPs open

Custom MCPs looked like a product benefit until they became a compute extraction method.

  • Claire saw a custom MCP behave like a nonstop worker running through ChatPRD.
  • The better approach is to restrict suspicious hosts, unsupported tunnels, and untrusted MCP setups.

Practical Application

Audit every free path

List every way a non-paying or lightly paying user can access model inference.

  • Include free trials, promo codes, referral credits, support workflows, onboarding demos, and custom integrations.
  • For each path, estimate the worst-case model bill, not just the average use case.
  • Kill or constrain any path where the downside is unbounded.

Add a model gateway

Put all LLM calls through a gateway or central routing layer.

  • Define an explicit allowlist of supported models.
  • Block direct calls to premium models unless the product deliberately exposes them.
  • Add spend caps and alerts per account, per organization, and globally.

This gives PMs and engineers one place to manage cost risk instead of scattering model logic across the app.

Review spend every hour

Set up an operational cadence that treats model spend like a security signal.

  • Monitor spend at least hourly if real-time alerting is not ready.
  • Pair spend spikes with account metadata, country, model, and payment signals.
  • Create a clear response rule for what happens when thresholds are crossed.

Claire's lesson is that waiting until the end of the billing cycle is far too late.

Build a fraud classifier

Create a lightweight AI triage loop for suspicious activity.

  • Feed in chat metadata, signup patterns, payment signals, and abuse thresholds.
  • Classify likely cases such as trial abuse, promo abuse, card fraud, resale, or legitimate heavy use.
  • Route the result to one of three actions: manual review, cheaper model, or account ban.

This is a realistic project for a PM to scope with engineering because it is mostly workflow and policy, not frontier research.

Lock down support actions

Review what your support AI or human support team can do after a ban.

  • Remove any easy path to unban, reissue promo codes, or extend free access without review.
  • Give support playbooks for the two common post-ban cases Claire saw: garbled scam appeals and polished but invalid appeals.
  • Treat support tooling as part of the abuse surface, not just a service layer.

Questions to Consider

Where are our zero-dollar doors?

Which free trials, promo codes, onboarding flows, or support paths in your product currently provide access to expensive model calls without a hard cost ceiling?

What does one bad actor cost?

If one determined user targeted your most expensive model through the easiest signup path, how much could that user cost in one hour, one day, and one week?

Which flexible features are risky?

Which product features in your roadmap, such as custom integrations, MCP support, or generous self-serve setup, increase user value while also creating new abuse surfaces?

Can our support desk be exploited?

Could a banned user regain model access today by persuading a support agent, a support chatbot, or an automated billing workflow to reverse a decision?

Bottom Line

Claire Vo's lesson is simple and expensive: an AI product is a compute business whether you price it that way or not. PMs building AI products need to design trials, pricing, support, model routing, and extensibility around abuse resistance from day one, because attackers will find every cheap path to premium inference.

Resources Mentioned

ResourceTypeWhy it was mentioned
Claire Vo on token fraud at ChatPRDX threadClaire says this post triggered a wave of responses from founders and operators dealing with the same abuse patterns.
Gareth Wham on fighting AI abuseX threadClaire cites Gareth's list of business controls as a useful reference for how larger companies handle similar problems.
FBI Internet Crime Complaint Center (IC3)Government resourceClaire says she submitted cybercrime reports for what looked like coordinated international abuse efforts.
Simon Willison's pelican riding a bicycleBlog postClaire references the well-known image prompt while describing strange probing and image-generation behavior she saw in abused accounts.
How I AI: I'm using Jev more than Opus 5.5 or GPT-6Podcast episodeThe feed links this earlier episode because Claire relies on Jev heavily for low-cost classification and decision loops.

Concepts to Explore

Adversarial unit economics

This episode is really about adversarial unit economics: what happens when bad actors optimize against your cost structure harder than good users do. PMs building AI products should study this because normal average-case pricing models can fail once attackers target the most expensive model paths.

Client attestation

Claire's point about verifying the client, not just the login, points toward client attestation and proof-of-origin patterns. These matter when a valid credential can still be used through relays, automation, or reverse proxies that the product never intended to support.

Abuse-resistant PLG design

Product-led growth usually assumes low-cost self-serve access. AI products challenge that assumption because self-serve access can expose expensive compute immediately. PMs should explore how PLG changes when variable cost is high from the first user action.

Case Studies Mentioned

ChatPRD free trial farming

Claire says free trials were repeatedly used by non-customers to access expensive models for unrelated work. The lesson was not just that abuse exists. It was that the economics of the free tier no longer worked, which pushed her to eliminate free trials entirely.

Promo code resale loops

Promo codes meant to help partners and users became a resale channel. Claire says some codes were sold on Reddit-style reseller markets and reused to spin up other AI businesses. The product lesson is that goodwill-based distribution mechanisms do not survive once they unlock valuable compute.

Reverse proxy resale scheme

Claire describes a case where another AI product authenticated into ChatPRD through a web browser, relayed requests to ChatPRD, then resold the outputs to its own users. The lesson is that your competitor or reseller does not need direct API access if your product surface is open enough to be proxied.

Custom MCP worker abuse

A user installed a custom MCP into ChatPRD that effectively ran as a nonstop compute worker. That changed Claire's view of MCP support from a feature to monitor into an abuse vector to constrain.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
Stripe RadarPayment fraud detection and rule engineClaire uses it heavily for card fraud, geography checks, and custom rules such as spotting repeated use of the same card across multiple trials.
Vercel AI GatewayCentral routing layer for model callsClaire uses it to control spend and restrict which LLM models ChatPRD will actually allow.
Vercel Bot IDBot verification for app requestsClaire uses it to confirm that traffic is really coming from the product, not just from stolen credentials or scripts.
Vercel FirewallNetwork-level traffic filteringClaire uses it for country blocks, VPN controls, IP blocking, and rate limiting.
ClerkAuthentication providerClaire says auth providers often have bot-protection settings that teams should explicitly turn on and tune.
CodexAI coding assistantClaire used Codex to help write Stripe Radar rules and even to help author FBI cybercrime reports.
JevDecision model and workflow toolClaire uses Jev to classify suspicious activity into fraud categories and trigger bans or model downgrades.

People to Follow

Claire Vo

Claire Vo is the founder of ChatPRD and host of How I AI. She is useful to follow because she shares operator-level lessons from running a bootstrapped AI product, including the messy economics and abuse problems many teams avoid discussing publicly.

Gareth Wham

Gareth Wham from Atlassian is mentioned as someone who publicly shared practical controls for fighting AI abuse. Claire points to his post as evidence that token fraud is not just a small-company problem.

Notable Quotes

"If you sell AI, you sell compute." (Claire Vo)

"Your own AI will help them if you let it." (Claire Vo)

"You have to verify your client, not just the login." (Claire Vo)

"Every zero dollar door gets used every trial, every coupon code." (Claire Vo)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free