Definition
Phishing works because the recipient has no way to check "did this email really come from that company?" The sender's display name, logo, wording, and the look of the linked page can all be copied. Sender domain authentication (SPF, DKIM, DMARC) lets receivers reject some spoofed mail, but not mail sent from a confusingly similar domain.
Personal data stolen through unauthorized access makes phishing more precise. When a name, registered address, or past purchase matches reality, the recipient is more likely to believe it. In Japan, a 2012 amendment to the Act on the Prohibition of Unauthorized Computer Access made setting up phishing sites, and obtaining others' identification codes through them, punishable.
People aren't the only ones fooled. In a Varonis test, an AI agent that received an urgent-looking request email sent (simulated) credentials outside the company before verifying the sender — even though it had been configured to verify the requester first (Critical Brief 047).
Lemma implementation
As long as authenticity is judged by appearance, fakes can be made. Lemma makes the origin of messages and documents cryptographically checkable. The issuer signs the content, and the receiver verifies "this issuer sent this content" through a provenance proof, not by how it looks.
Where legitimate senders attach proofs to their requests, AI agents can be given a rule to check the origin proof of incoming requests before acting on them. Requests without a proof, or whose verification fails, stop before reaching operations like sending credentials or making payments.
The record of checking the origin is kept as an audit trail as well. Afterwards, it can be traced which request was judged genuine, and on what basis.