Continuous adversarial testing for fraud defense stacks
A red-team service that attacks your fraud rules and models with simulated adaptive fraudsters, measuring how fast each defense gets bypassed before real attackers find out.
A ready-to-paste brief for Claude Code, Cursor or any coding agent.
Overall6.2/10
Pain7
Willingness to pay6
Timing6
Gap7
Moat5
Feasibility5
For
- Who pays
- Heads of fraud/risk at mid-market fintechs and marketplaces (500–5,000 employees), $3–10k/mo from the existing fraud tooling budget alongside Sift/Sardine spend.
- What people said · 3
- The post itself states that trained models and rulesets become obsolete within weeks as fraudsters reverse-engineer the detection logic, forcing constant retraining. From the article ↗
- A veteran of AOL's spam wars says every defensive technique they built over weeks or months was circumvented by adversaries in a single weekend, and attackers have only gotten faster since. Read the comment ↗
- A commenter dismisses surface-level heuristics like asking a model whether an email looks disposable, arguing real fraud signals run much deeper than appearances — implying defenses are validated far too shallowly. Read the comment ↗
Paraphrased from the discussion. Follow the links for the original comments.
- Competitors and gap
All existing fraud platforms, including Kairos, are detection engines: they score and decide on live traffic. None continuously probes the customer's own ruleset and models with mutated, adversarial identities and behaviors to report time-to-bypass and decay curves before real fraudsters exploit the holes. Pen-testing firms do this manually for appsec, not for fraud decisioning logic.
- Why now
- LLM agents make it newly cheap to generate large volumes of plausible adversarial behavior (varied identities, devices, narratives, session patterns), and the shift to agent-tuned fraud stacks (as Kairos shows) creates machine-readable rulesets and models that can be attacked programmatically via the same APIs.
- Moat · 5/10
- Compounding advantage: every engagement generates a proprietary corpus of bypass techniques and decay benchmarks across customers, which improves the attack generator and becomes a defensible dataset no detection vendor has.
Against
- Red team
- Detection vendors (Sift, Sardine) genuinely don't continuously attack the customer's own rulesets to measure time-to-bypass, and pen-testing firms don't touch fraud decisioning logic, leaving a real gap with a buyer who already spends on fraud tooling.
- What would kill it
- Fraud teams treat this as a nice-to-have audit rather than a recurring need, and the detection vendors (Sift, Sardine, Kairos) add basic simulation as a feature, collapsing the standalone market.
Start
- First three steps
- Build an attack generator that mutates identity/device/session/event payloads and replays them against a customer's ruleset via API to find bypasses.
- Ship a rule-decay report: replay historical confirmed-fraud events weekly and quantify how detection recall degrades over time per rule/model.
- Deliver a prioritized fix list (which rule, which variant, expected lift) that fraud engineers can hand to their decisioning platform.
- Source discussion
- Show HN: Kairos – Open-source risk and fraud decisioning platform ↗ 103 points30 commentsFrom the storyLinked article