AI-powered account attacks
Attackers use AI to make the old attacks more convincing and cheaper to run at scale. The FBI warned in December 2024 that criminals use generative AI to write believable messages and build fake profiles and websites in volume, to generate fake identity documents and photos, and to clone the voice of a relative or an official from a short clip so a call sounds real.
For accounts, that adds up to three main threats:
- Phishing that looks right. Well-written emails and pixel-perfect login pages, tailored to you, that collect a password and the one-time code you type next.
- Voice and video impersonation. A call that sounds like your manager, your bank or a family member, asking you to approve a payment, read out a code or reset access.
- Password attacks at scale. Automated tools that try passwords leaked from other sites against your accounts, and keep trying.
All three depend on you handing over something reusable: a password, a code or an approval. The best defences remove that something.
Phishing-resistant sign-in with passkeys
The strongest protection for an account is sign-in that cannot be phished. CISA’s guidance calls phishing-resistant MFA, meaning FIDO or WebAuthn authenticators such as passkeys and hardware security keys, the gold standard, and urges organisations to make migrating to it a high priority.
Passkeys resist phishing because of how they work. Your device holds a private key, the site holds the matching public key, and at sign-in the device signs a challenge from that exact site. The private key never leaves your device, and the signature is bound to the real site’s address, so a look-alike page cannot get anything it could replay. There is no code to read out and no password to type into the wrong box.
Any second factor is far better than none. Microsoft reported that MFA can block over 99.9 percent of account compromise attacks. But CISA notes that weaker forms can be beaten: codes sent by text are exposed to SIM swaps and phone network attacks, one-time codes can be typed into a phishing page, and push approvals can be spammed until someone taps Accept. Turn on passkeys or a security key wherever an account offers them, starting with email, because whoever controls your email can reset everything else.
NIST password guidelines
Every password should be long, unique and stored in a password manager. NIST’s current guidelines require at least 15 characters for a password used on its own, tell services to check new passwords against lists of commonly used, expected or compromised values, and say services shall not force periodic changes, only a change when there is evidence of compromise. A unique password per account means a leak at one site does not open the others.
AI voice and message scam prevention
The defence against AI voice and message scams is to verify through a channel the attacker does not control. The FBI’s advice is to hang up, look up the organisation’s contact details yourself and call the number directly, and to agree a secret word or phrase with family that a caller has to know.
- Treat urgency as a warning sign. Requests to act now, keep it quiet or skip the usual process are how these scams work.
- Never read out a one-time code or approve a sign-in you did not start.
- Lock down account recovery. Use recovery codes or a second passkey rather than SMS, and ask your mobile carrier about a port-out or SIM lock.
Device-held keys for account security
Device-held keys leave nothing to steal from the conversation. A password, a code or a bearer token is a secret that works for whoever holds it, so a convincing fake only has to get you to hand it over once. A key that stays on your device and proves itself by signing a fresh challenge gives a fake page or voice nothing reusable.
Offline Protocol applies the same principle to devices. Every device generates its own Ed25519 identity key and keeps the private half; its address is derived from the public key, so other devices verify it directly with no lookup and no shared secret. When an app binds a device to an account, the server issues a fresh nonce, the device signs it, and the server verifies the signature. Nothing a phishing page or a cloned voice could capture would let anyone sign as that device.