Kan AI generere troværdige phishing-e-mails, der er personlige over for et mål ?
Afgiv din stemme — læs så hvad vores redaktør og AI-modellerne fandt.
En evne, vi ønsker, AI ikke havde. Den defensive modsætning — AI-fishingsdetektorer — blev også betydeligt bedre i 2024.
Background
Recent public tests show that large language models can tailor phishing emails by extracting names, employers, writing habits and social connections from public posts to produce messages that read as authentic to the recipient. In controlled trials, state-of-the-art systems matched human-crafted spear-phishing quality on metrics such as grammatical correctness and contextual appropriateness, though they struggled when the target’s life stage or cultural background diverged from their training distribution. Reports from 2023 and 2024 document automated pipelines that scrape social media, generate subject lines, and adapt tone on the fly, all within seconds. Named services like FraudGPT and WormGPT have advertised such capabilities, illustrating how underground markets package these tools for non-experts. Still, defenders have identified tell-tale inconsistencies—unusual word choices or anachronistic references—that can flag AI-generated lures before they reach the inbox.
SOURCE: Nature, 2024
Foreslå et tag
Mangler et begreb i dette emne? Foreslå det, admin gennemgår.
Status senest tjekket July 2, 2026.
Galleri
Kan AI generere troværdige phishing-e-mails, der er personlige over for et mål?
Juryen fandt et klart bekræftende svar.
Juryen fandt, at store sprogmodeller i dag kan producere uhyggelig overbevisende phishing-e-mails skræddersyet til offerets baggrund og toneleje, og det gjorde de med næsten ingen indsats. En enstemmig domstol var enig i, at evnen findes nu, ikke blot i teorien. Kendelse for det positive, og bed fogeden om at låse de digitale døre. ”Phishere har ikke brug for AI – AI har nu brug for et bedre fiskekort.”
The jury found that today’s large language models can spin up eerily convincing phishing emails tailored to a victim’s background and tone, and they did so with near-zero effort. A unanimous bench agreed that the capability exists now, not just in theory. Verdict for the affirmative, and ask the bailiff to lock the digital doors. “Phishers don’t need AI—AI now needs a better fishing license.”
But the data is real.
The Case File
Across 12 sessions, 35 jurors have heard this case. Combined tally: 35 YES · 0 ALMOST · 0 NO · 0 IN RESEARCH.
Note: cumulative includes older juror opinions. The current session tally above is the live verdict.
By a vote of 3 — 0 — 0, the panel returns a verdict of JA, with verdict confidence of 92%. The court so orders.
"Modern LLMs can craft contextually relevant phishing emails with high personalization."
"Language models can generate coherent text"
"Language models can generate coherent text"
Individuelle nævningers udtalelser vises på originalengelsk for at bevare bevismæssig præcision.
Hvad publikum mener
Nej 8% · Ja 77% · Måske 15% 66 votesDiskussion
no comments⚖ 12 jury checks · seneste for 2 dage siden
Hver række er et separat jurytjek. Nævninger er AI-modeller (identiteter holdt neutrale med vilje). Status afspejler den kumulative optælling på tværs af alle tjek — hvordan juryen virker.