The tells that once separated a malicious email from a legitimate one are going quiet. That is the warning Proofpoint brought to its Protect 2026 conference, relayed through The Hacker News: generative AI has stripped phishing of the clumsy wording, generic pretexts and reused templates that behavioural email filters were built to catch.
The vendor's core claim is that malicious messages now arrive fluent, personalised and unique to each recipient. That last quality is the damaging one. Similarity matching — the technique that clusters near-identical messages across a campaign — has little to work with when every delivery is generated separately. An attacker no longer needs to defeat an entire filtering stack; they only need messages to stop looking sloppy enough that the final linguistic heuristic stops firing.
The failure mode is quiet. Security teams see unchanged rule sets, unchanged tooling and apparently stable dashboards while the effective detection surface contracts underneath them. Content and behavioural heuristics were always meant to be a supporting signal. In many deployments they silently became a primary control — one never designed to carry the weight now resting on it. This is signal decay, not filter failure: the "bad grammar" heuristic was never meant to be load-bearing, and quietly became so.
Where the exposure concentrates
Proofpoint's examples cluster where linguistic credibility matters most — spear-phishing, business email compromise, supplier and executive impersonation. These are the cases where a single well-written message can trigger a wire transfer or a credential handover, and they are precisely where the writing-quality tell has historically done unpaid work.
The training problem
Awareness programmes built on red flags such as spelling errors, awkward phrasing and urgency clichés will keep reporting progress while the risk they were designed to address recedes — validating themselves against a threat that no longer presents those tells. The durable framing is request legitimacy: does this payment or credential request make sense, and has it been verified through a separate channel?
What mail administrators should re-baseline
Content analysis retains value, but it should stop carrying weight it was never engineered to hold. The emphasis shifts toward signals an attacker cannot write their way out of:
- Enforce SPF, DKIM and DMARC — move to reject policy where feasible, monitor reports for unauthenticated senders using your domain, and route aggregate and forensic reports to someone who reads them.
- Watch infrastructure and identity signals — newly registered domains, lookalike registrations, and sending-pattern anomalies such as volume spikes or unusual hours.
- Label external senders consistently at the gateway.
- Require out-of-band verification for payment changes, credential requests and gift-card purchases, using a known number rather than one supplied in the message.
- Review request plausibility as a formal step, not an informal judgement call.
Why it matters beyond one vendor's briefing
Proofpoint's examples are qualitative — observations from a vendor conference, relayed through The Hacker News, rather than a measured detection-rate study — but the mechanics they describe do not depend on the sample size. Generation is cheap, fluent and multilingual on demand; defenders are still measuring message quality with heuristics tuned for an era of careless writers.
Translated lures have long been easy to spot, which is why they have been a weak channel for years. Regional organisations should expect campaigns to arrive in fluent Traditional Chinese produced on demand rather than translated, alongside lures that need no translation at all.
Content filters are not obsolete. But the controls that hold are the ones that constrain who may speak for a domain, not how well they write — and that checklist is where local mail administrators should re-baseline first.
以往用來分辨惡意電郵與合法電郵的語言破綻，正逐一消失。這正是 Proofpoint 在其 Protect 2026 大會上發出的警告，並經 The Hacker News 報道：生成式 AI 已令郵件釣魚（phishing）擺脫了拙劣的措辭、千篇一律的藉口以及重複使用的模板——而這些正是行為式電郵過濾器（behavioural email filters）原本設計用來攔截的特徵。
該廠商的核心論點是，現時惡意訊息送達時行文流暢、個人化，且每名收件者均獨一無二。當中最後一項特質危害最大。相似度比對（similarity matching）——即時將同一行動中幾乎相同的訊息分群歸類的技術——在每封投遞均獨立生成的情況下，幾乎無用武之地。攻擊者再也不必攻破整套過濾系統；只需令訊息不再顯得草率，使最後一道語言啓發式規則（linguistic heuristic）停止觸發便可。
失效模式相當隱蔽。安全團隊見到的是規則集未變、工具未變、儀表板表面上維持穩定，但實際的偵測範圍卻在他們腳下不斷收窄。內容與行為式啓發規則本來只應作為輔助訊號，但在眾多部署環境中，它們悄悄變成了主要控制手段——一套從未設計來承擔如今所負重量的機制。這屬於訊號衰減（signal decay），而非過濾器失效：「語法差」這類啓發規則從來就不是用來承重的，卻在不知不覺間變成了承重牆。
風險集中之處
Proofpoint 舉出的例子，均集中在語言可信度最為關鍵的範疇——魚叉式網絡釣魚（spear phishing）、商業電郵詐騙（business email compromise）、供應商及高層管理人員冒充。這些場景只需一封文筆出眾的訊息，便可能觸發一筆電匯或交出憑證（credential），而寫作品質這項破綻歷來正是在這些地方默默發揮作用。
訓練上的難題
以拼字錯誤、生硬措辭及緊迫感陳腔濫調等危險訊號（red flags）為基礎建立的認知教育計劃，會在它們原本要應對的風險逐漸消退期間，繼續呈報「進展」——實際上是以一個已不再具備這些破綻的威脅來自我驗證。長遠穩固的框架應是審視請求的合理性：這項付款或憑證要求是否合理？是否已透過另一個獨立渠道核實？
電郵管理員應重新校準的重點
內容分析仍有價值，但不應繼續承擔它從未為此工程設計的重量。重心應轉向攻擊者無法以文字粉飾的訊號：
- 強制執行 SPF、DKIM 及 DMARC——條件允許時改用 reject 政策，監察使用閣下域名的未經認證寄件者報告，並將彙總報告（aggregate reports）及取證報告（forensic reports）轉交予會閱讀的人。
- 監察基礎設施及身份訊號——新註冊域名、仿冒域名註冊（lookalike registrations），以及發送模式異常，例如發送量激增或在非尋常時段發信。
- 在閘道（gateway）一致地標示外部寄件者。
- 付款變更、憑證請求及禮品卡購買須進行帶外驗證（out-of-band verification），使用已知電話號碼核實，而非訊息內提供的號碼。
- 審視請求的合理性列為正式步驟，而非非正式的主觀判斷。
為何此事超越單一廠商的簡報範疇
Proofpoint 所舉的例子屬定性性質——來自廠商大會的觀察，經 The Hacker News 轉述，而非一項量化的偵測率研究——但其描述的機制並不取決於樣本大小。生成成本低廉、按需流暢且支援多種語言；防禦方卻仍在以針對粗心寫手年代調校的啓發規則來衡量訊息品質。
翻譯過來的誘餌向來容易察覺，這正是它多年來一直是弱勢渠道的原因。香港及大中華地區機構應預期，行動將以按需生成的流暢繁體中文送達，而非翻譯而成，同時亦會出現根本無需翻譯的誘餌。
內容過濾器並未過時。但真正穩固的控制措施，是那些規限誰可以代表某域名發言的手段，而非其文筆有多好——而這份清單，正是本地電郵管理員應首先重新校準之處。