Binance Reports AI-Powered Anti-Scam Progress: Over $10.53B in Losses Prevented and 22.9M Attacks Blocked
AI has changed the crypto security game in two directions at once: it helps defenders detect fraud faster, and it helps attackers scale scams with deepfakes, voice cloning, phishing bots, and impersonation at a level that looks “human” to both users and basic filters.
A recent analysis from Binance Research details what this arms race looks like at exchange scale — and why everyday users still need strong self-custody and operational security habits to stay safe in 2026. You can read the full piece here: AI-Powered Crypto Security.
The headline numbers (and why they matter)
Crypto security headlines often focus on smart contract exploits. But in practice, the fastest-growing losses still come from social engineering: convincing people to approve transfers, reveal credentials, or “verify” a wallet.
According to Binance Research (data timestamped as of April 29, 2026), Binance’s AI-assisted defenses reported the following results:
- FY 2025 (ending November 2025): blocked $6.69B in fraud and scam attempts across 5.4 million users, blacklisted 36,000 addresses, and issued 9,600+ real-time pop-up warnings daily.
- Q1 2026: intercepted 22.9 million scam and phishing attempts, safeguarding $1.98B in user funds.
- Cumulative impact (2025 → Q1 2026): $10.53B in potential user losses prevented.
These numbers are important for the industry narrative: centralized platforms are increasingly being judged not only by liquidity and uptime, but by their ability to operate as active fraud-prevention networks — in real time, at massive scale. (binance.com)
Why AI-driven scams are harder to spot than “classic” crypto fraud
The old scam playbook was relatively easy to pattern-match: poor grammar, obvious copy-paste templates, and low-effort fake websites. AI flips that.
Today’s scams are increasingly:
- Hyper-personalized (built from breached data and social profiles)
- Multi-channel (SMS + email + Telegram/WhatsApp + phone calls)
- Real-time adaptive (chatbots that change tactics based on your replies)
- Audio/visual convincing (voice cloning and deepfake video)
This is not just a crypto problem. Reports on malicious AI use cases show that threat actors commonly combine AI with “traditional” infrastructure like social accounts, websites, and scripted outreach — but AI makes the workflow cheaper and easier to scale. See: OpenAI’s report on disrupting malicious uses of AI. (openai.com)
Binance’s defense approach: models, monitoring, and “friction at the right moment”
Binance Research describes a layered approach that focuses on behavioral signals rather than only static rules.
1) AI models and compliance initiatives
Binance reports 24+ AI initiatives across compliance and 100+ AI models powering anti-fraud controls, alongside a claim of materially reducing illicit fund exposure. (binance.com)
2) Behavior-based detection (especially around withdrawals)
Even if platforms cannot stop every phishing attempt upstream, they can reduce damage by detecting anomalies such as:
- unusual login locations
- repeated failed attempts
- abnormal payment patterns
- sudden large withdrawals
This is where “AI fraud detection” becomes tangible: it introduces timely friction — warnings, additional verification steps, delayed withdrawals, or manual review — right when a scammer is pressuring the victim to act fast. (binance.com)
3) Simulating phishing environments to reduce success rates
Binance Research also cites a phishing simulation technique that reduced the phishing rate from 3.2% to 0.4% (an 8x improvement). (binance.com)
The broader trend: scam losses keep rising across the ecosystem
Even with stronger exchange-side defenses, the overall scam economy remains huge.
- Chainalysis estimates that $17B was stolen in crypto scams and fraud in 2025, driven in part by a sharp rise in impersonation scams. Read the relevant analysis here: Chainalysis on 2025 scam and fraud activity. (chainalysis.com)
- In the United States, the FBI’s IC3 report shows $11.366B in losses tied to cryptocurrency complaints in 2025 (with 181,565 complaints). See the source document: 2025 IC3 Annual Report (PDF). (ic3.gov)
The takeaway: security is improving, but attack volume and scam quality are improving too. The user is still the most targeted surface area.
What users should do in 2026: a practical anti-scam checklist
Below is a high-signal checklist focused on crypto scam prevention, account takeover defense, and self-custody safety.
1) Harden exchange accounts against phishing and SIM swaps
- Use app-based 2FA (or passkeys, where available) instead of SMS.
- Enable an anti-phishing code (so real emails are easier to distinguish from fakes).
- Turn on withdrawal address allowlisting and time locks where supported.
- Treat “support calls” and “urgent security checks” as hostile by default — end the conversation and verify using official channels.
2) Assume deepfakes will be used against you
Deepfake audio/video is most effective when the attacker creates urgency:
- “Your account is compromised.”
- “Funds must be moved to a safe wallet.”
- “You must verify by sharing a code / seed phrase.”
A safe rule: no legitimate support process ever requires your seed phrase. Not now, not later.
3) Limit blast radius when using AI trading tools, bots, or plugins
Agentic tools are powerful — and that power is exactly why they need strict permissions.
Binance highlights an important design principle in its discussion of Binance Ai Pro: fund segregation plus no withdrawal permissions for the agent. (binance.com)
If you want more context on the product’s structure, see: Binance Ai Pro guide. (academy.binance.com)
Regardless of platform, apply these rules:
- Use a separate sub-account or limited-balance account for automation.
- Do not grant withdrawal-enabled API access to third-party tools.
- Avoid “installing plugins” from unverified sources. If a platform blocks a plugin for risk reasons, treat that as a signal, not an inconvenience.
4) Move long-term holdings to cold storage
Exchange security can reduce losses, but it cannot eliminate user-side compromise. For long-term holdings, cold storage remains one of the cleanest ways to reduce phishing-driven loss, because private keys never touch an internet-connected environment.
This is where a hardware wallet fits naturally into a modern security stack:
- It keeps private keys offline
- It forces on-device confirmation for transactions
- It reduces the chance that a compromised laptop/phone can silently approve transfers
If you’re building a self-custody setup, OneKey is designed around these fundamentals, and also supports advanced protections like passphrase (hidden wallets) — helpful for reducing single-point-of-failure risk when seed backup hygiene is strong.
Final thoughts: “AI security” is not a product — it’s a posture
The most important signal from Binance’s numbers is not just the dollar amount prevented; it’s the direction of travel: crypto platforms are becoming AI security operators, while attackers are becoming AI-enabled social engineers.
In that world, the winning strategy for users is layered:
- strong account security + anti-phishing habits
- least-privilege automation
- cold storage for long-term assets
- relentless skepticism toward urgency and impersonation
If you treat security as an ongoing process rather than a one-time setup, you’ll be aligned with where the crypto industry is heading in 2026 and beyond.
Further reading
