AI can make scam messages and voices convincing. Verify the requested action through an independent channel, especially before sending money or approving access.
Verify independently; scan the device after suspected malicious downloads. See FTC phishing guidance.
The defensive answer is not "spot the typo". It is to protect your control plane and to treat money, access, and verification codes as high-risk events that always require a second channel.
Decisions that stop most AI-assisted scams
| Request or signal | Safe default | Why it works |
|---|---|---|
| Payment change, new bank details, gift cards, crypto, or urgent invoice | Verify out of band (call a known number, not the number in the message). | AI makes messages believable, but it cannot easily defeat a second trusted channel. |
| "Approve this login" prompts or repeated MFA notifications | Assume an active takeover attempt and contain immediately. | Many compromises succeed when victims get tired and approve one prompt. |
| Login link, QR code, or "security alert" in email or DMs | Navigate to the site directly, then check real account notifications. | Phishing works by controlling the click path. |
| Someone asks for a code, screenshot, or "verification" | Do not share codes. Treat it as a takeover attempt. | Codes and screenshots are often enough to reset accounts or bypass checks. |
Rule of thumb: if a message can move money or change access, treat it as hostile until verified.
What AI actually changes: and what it does not
AI does not create new human weaknesses. It industrializes old ones.
- Language quality: clean grammar is no longer a sign the message is legitimate.
- Personalization: attackers can stitch together public data into believable context fast.
- Volume: more outreach means more chances to hit someone at the wrong moment.
- Impersonation: voice and video deepfakes increase pressure to comply, especially in urgent scenarios.
What does not change is the control-plane reality: most takeovers still succeed by stealing credentials, stealing sessions, or abusing recovery paths. If your inbox and recovery methods are weak, AI is not the problem. It is just the multiplier.
Common AI-assisted attack patterns
Polished phishing that looks like normal business
Attackers can generate emails that match the tone of a coworker, a vendor, or a platform notification. That makes social engineering harder to spot in isolation. The correct response is to verify the action, not to debate the writing style.
Deepfake voice or video pressure
Deepfakes are most effective when paired with urgency: "approve this", "send this", "do this before the deadline". The defense is pre-commitment. Decide in advance that certain actions require a second channel, even when the request seems to come from a familiar face or voice. For background on the failure modes, see deepfakes and why they are dangerous.
Fake support and recovery scams
When people are locked out or see an outage, scammers flood search results and DMs with fake support. They try to get remote access, codes, or payment. The safe pattern is to use official, in-app recovery routes and to ignore unsolicited "support" contact.
QR code phishing: quishing and link laundering
AI is often used upstream to generate convincing pretexts for QR codes and links. If you scan a QR code, verify the destination before you log in. Use QR code phishing (quishing) for practical checks and containment steps.
MFA prompt abuse: push bombing
Repeated prompts are a sign someone is trying to force an approval. Do not wait for it to stop. Use MFA fatigue (push bombing) for the containment sequence.
Defensive controls that still work
Protect the inbox first
Email is the reset button for most services. If an attacker gets your inbox, they can reset other accounts, intercept recovery links, and lock you out. Secure the inbox with strong authentication and session review before you focus on any single platform.
Use phishing-resistant authentication where possible
Passwords get stolen and reused. Strong factors reduce the value of a stolen password, and phishing-resistant methods reduce the value of a fake login page. If you want the concept overview, see passkeys and two-factor authentication (2FA).
Stop relying on "looks real" as a test
Train yourself to look for decision triggers instead of style triggers. A message is high risk if it requests:
- money or payment changes
- codes, screenshots, or "verification"
- software installs, browser extensions, or remote access
- access changes: new admins, new devices, recovery email changes
Use how to identify scam emails for a more complete pattern set.
Organizations: tighten email authentication and reduce spoofing surface
AI increases the effectiveness of impersonation, but basic email authentication still matters. If you run a domain, implement and maintain SPF and DKIM so spoofed mail is easier to filter and investigate. See Sender Policy Framework (SPF) and DomainKeys Identified Mail (DKIM).
Teach verification rituals, not fear
Training works when it gives employees a small number of repeatable actions. The best ones are out-of-band verification and refusal to share codes. For a focused workflow, use train employees to spot phishing emails.
Evidence to keep for a scam report
Keep the claimed identity, requested action, independent verification result, and any payment or account access granted. Record dates, times, and what you actually did. Separate confirmed events from the other person's claims, and retain original files alongside screenshots.
After urgent account or payment action, you can organize the evidence in a Scam Incident Record. Keep the relevant messages and transaction details together for reporting and follow-up. You review and send information to the appropriate organization yourself; creating a record does not submit an official report or guarantee recovery.
Leave passwords, login codes, backup codes, and private keys out of incident notes or reports. Submit identity documents only when the verified official recovery process requires them.
What to do if you clicked or responded
If you entered credentials, approved a prompt, shared a code, or installed something, treat it as compromise until proven otherwise. Contain first:
- secure the email inbox and phone recovery channels
- end active sessions and revoke unknown access where possible
- reset passwords to unique values stored in a password manager
- enable stronger authentication and review recovery methods
If the incident involved a suspicious download or browser extension, assume device compromise. Start with infostealer malware to understand why password changes alone can fail.
AI-assisted phishing will keep evolving because it is profitable. The way to win is to make the scam unreliable. When verification is mandatory for money and access changes, and when your recovery channels are hardened, the attacker’s best messages still do not produce outcomes.
Over time, the strongest posture is simple: protect the inbox, use stronger authentication, and verify high-risk requests out of band. That is how you reduce risk even when the messages get better.
The practical goal is not to recognize every scam. It is to ensure that a single moment of confusion cannot move money, change access, or take over the accounts that control everything else.