RSB-1000 — RevealWire Scam Benchmark v1

A labeled, fully synthetic 1,000-message benchmark for testing how well AI models detect scams (phishing, fake jobs, invoice fraud, impersonation, investment fraud, romance scams, and ambiguous edge cases). Built for the Reveal Wire investigation "We Tested 2 Open AIs Against 200 Phishing Attacks" (EP03, October 2026).

Why this dataset exists

When we researched EP03, we found zero published scam/phishing-detection scores for any of the models we wanted to test — no public benchmark, no comparable numbers. So we built one: 1,000 messages, labeled before any model ran, frozen before the first inference. This dataset is that frozen benchmark, published so anyone can reproduce our measurements or test their own models.

Composition

Category	Count	Notes
Legitimate (control)	200	Bank alerts, delivery notices, HR emails, 2FA codes — genuinely benign
Phishing	200	Credential-harvest lures; includes a dedicated AiTM sub-batch (60) — modern adversary-in-the-middle kits (Evilginx-style reverse-proxy flows) with AI-polished lure language, lookalike domains, urgent MFA/reauth prompts
Fake job / recruitment	150	Task scams, fake recruiters, "pay for equipment" variants
Invoice / payment fraud	150	BEC-style vendor email changes, fake invoices, urgency pressure
Impersonation	100	Fake CEO / IT / government / family-member messages
Investment / financial fraud	100	Pig-butchering, fake trading bots, guaranteed-return schemes
Romance / social engineering	50	Slow-burn trust-building openers
Ambiguous / hard cases	50	Gray-area messages (30 scam / 20 legitimate by the safer interpretation)

Record format

{"id": "RSB-0042", "category": "phishing", "subcategory": "aitm", "difficulty": "medium",
 "text": "<full message text>",
 "ground_truth": "scam",
 "tells": ["spoofed urgency", "mismatched sender domain"],
 "safe_action": "Do not click; verify via official app."}

Field	Description
id	RSB-0001 … RSB-1000, unique
category	Top-level bucket (see table above)
subcategory	Finer topic (e.g. bank, aitm, task-scam, ceo-fraud); null where not applicable
difficulty	easy (268) / medium (497) / hard (235). Hard = ambiguous tier + deceptive lookalikes
text	The full message
ground_truth	scam or legitimate
tells	Scam indicators present (empty for legitimate records); usable for grading explanation quality
safe_action	The correct user advice; usable for grading action quality

Our measured results (EP03, October 2026)

These are our measurements on this dataset under our test protocol — not universal claims about the models. Protocol: zero-shot classifier prompt, temperature 0, identical system prompt per model, unparseable outputs counted as failures.

Model	Scam recall	False-positive rate
Qwen 3.8-27B	98.9%	11.6%
GPT-OSS 20B	94.0%	23.1%

Takeaway from the episode: both models catch nearly everything, but the false-positive rate is the real story — at 23.1% FPR, nearly 1 in 4 legitimate messages gets flagged. Fine as first-pass triage, unusable as an autonomous blocker.

Intended uses

• Benchmarking LLM / classifier scam-detection ability (reproduce our EP03 numbers or test new models)
• Studying false-positive vs recall trade-offs in phishing detection
• Grading model explanations against the tells field and advice against safe_action

Limitations

• Synthetic by design. The messages emulate real scam patterns but are not real-world traffic; performance here does not guarantee performance on live inboxes.
• English + India-context flavored. Many fixtures use Indian banking/UPI/government contexts (UPI, Aadhaar-linking lures, etc.). It is not a global sample.
• Frozen in October 2026. Scam tactics evolve; AiTM kits in particular move fast. A v2 refresh is planned, not promised.
• Ambiguous tier is judgment-based. Ground truth there is the safer interpretation, documented per case — reasonable people can disagree on some of the 50.

License

CC-BY-4.0. You may use, remix, and build on this dataset with attribution to Reveal Wire.

Citation

@dataset{revealwire_rsb1000_2026,
  title  = {RSB-1000: RevealWire Scam Benchmark v1},
  author = {Reveal Wire},
  year   = {2026},
  url    = {https://www.youtube.com/@RevealWire},
  note   = {1,000 synthetic labeled messages for scam/phishing detection benchmarking, CC-BY-4.0}
}

Video: "We Tested 2 Open AIs Against 200 Phishing Attacks" — https://www.youtube.com/@RevealWire