Generative AI erodes behavioral email detection: Proofpoint details AI-crafted BEC, device-code phishing and real estate lure campaigns
Proofpoint reported at its Protect 2026 conference that generative AI is removing the behavioral tells (awkward language, reused templates, repeated infrastructure) that email filters have relied on. Attackers can now produce fluent, personalized, unique messages for each recipient, which weakens behavior-only detection. The vendor cites the FBI IC3 2025 Internet Crime Report: more than 22,000 AI-related complaints with nearly $900 million in losses, and 24,768 BEC complaints with about $3.05 billion in losses. Proofpoint argues defenders should shift from asking whether a message looks unusual to asking what it is trying to accomplish, and promotes its Nexus Intent-Based Detection Model and Threat Investigation Agent. This is largely a vendor-perspective report with product promotion, and no specific threat actor is named.
Technical Details
- Real estate campaign: one actor contacted 30+ agents at the same company from different real mailboxes, each with a different property enquiry, and ended every thread with a request to join a video meeting. In one thread, a 2:58 a.m. enquiry became a meeting invite 13 minutes later. The join link was a genuine google.com redirect whose URL parameter sent the click to an unrelated site, so reputation checks read clean. A variant targeting insurance agents used quote requests and referrals before a spoofed conference invite delivered the payload.
- Hostile communications: hundreds or thousands of senders wrote to the same executives with derogatory text or AI-generated or manipulated images. These carry no link or credential harvest and only show as a campaign when analyzed in aggregate, across text and imagery.
- Phishing-as-a-service kit: a single device-code phishing flow is presented behind four disguises (court access portal, document-signing request, identity check, finance report library). The page shows an attacker-generated code and sends the victim to a legitimate sign-in page to paste it, which authorizes the attacker's session.
- BEC example: an email with a fabricated quoted thread between an impersonated audit vendor and a target executive agreed an invoice, a 50% loyalty rebate and a payment date. It asked for same-day payment to a remittance address that differed from the sending address.
- Bank-detail change: Proofpoint lists five intent signals: the supplier's first banking change in years, a request that does not match the thread, a deadline tied to a payment run, a request to move off-thread to one person, and verification closed off with a claim that the finance mailbox is "mid-migration".
- Account takeover case study: an RFP-themed credential-phishing lure led to takeover of a finance user's account. The attacker signed in from New York and then Dallas, added a mailbox rule to hide replies, and sent about 100 phishing messages to 42 internal recipients. The agent contained it in four minutes and retrieved 92 outbound messages. Four higher-impact actions were held for human approval.
- Detection tiers: tier one under 500 ms for all mail, tier two for under 1% of messages, tier three for under 0.01%, with an organizational graph to correlate campaigns across mailboxes.
Remediation
- Keep signature, reputation and IOC-based detection and behavioral baselines, and add intent-based analysis of what a message is trying to get the recipient to do.
- Correlate similar requests across an organization and across mailboxes rather than scoring messages one at a time.
- Analyze text and images together to catch disruptive and harassment campaigns.
- Verify bank-detail changes and payment requests out of band, especially when the sender discourages verification or asks to move off the thread.
- Use automated investigation across email, identity, messaging and browser telemetry, with human approval for high-impact actions such as credential resets.
Linha do tempo
-
Generative AI erodes behavioral email detection: Proofpoint details AI-crafted BEC, device-code phishing and real estate lure campaigns
ler a análise desta fonte ocultar
Proofpoint reported at its Protect 2026 conference that generative AI is removing the behavioral tells (awkward language, reused templates, repeated infrastructure) that email filters have relied on. Attackers can now produce fluent, personalized, unique messages for each recipient, which weakens behavior-only detection. The vendor cites the FBI IC3 2025 Internet Crime Report: more than 22,000 AI-related complaints with nearly $900 million in losses, and 24,768 BEC complaints with about $3.05 billion in losses. Proofpoint argues defenders should shift from asking whether a message looks unusual to asking what it is trying to accomplish, and promotes its Nexus Intent-Based Detection Model and Threat Investigation Agent. This is largely a vendor-perspective report with product promotion, and no specific threat actor is named.
Technical Details
- Real estate campaign: one actor contacted 30+ agents at the same company from different real mailboxes, each with a different property enquiry, and ended every thread with a request to join a video meeting. In one thread, a 2:58 a.m. enquiry became a meeting invite 13 minutes later. The join link was a genuine google.com redirect whose URL parameter sent the click to an unrelated site, so reputation checks read clean. A variant targeting insurance agents used quote requests and referrals before a spoofed conference invite delivered the payload.
- Hostile communications: hundreds or thousands of senders wrote to the same executives with derogatory text or AI-generated or manipulated images. These carry no link or credential harvest and only show as a campaign when analyzed in aggregate, across text and imagery.
- Phishing-as-a-service kit: a single device-code phishing flow is presented behind four disguises (court access portal, document-signing request, identity check, finance report library). The page shows an attacker-generated code and sends the victim to a legitimate sign-in page to paste it, which authorizes the attacker's session.
- BEC example: an email with a fabricated quoted thread between an impersonated audit vendor and a target executive agreed an invoice, a 50% loyalty rebate and a payment date. It asked for same-day payment to a remittance address that differed from the sending address.
- Bank-detail change: Proofpoint lists five intent signals: the supplier's first banking change in years, a request that does not match the thread, a deadline tied to a payment run, a request to move off-thread to one person, and verification closed off with a claim that the finance mailbox is "mid-migration".
- Account takeover case study: an RFP-themed credential-phishing lure led to takeover of a finance user's account. The attacker signed in from New York and then Dallas, added a mailbox rule to hide replies, and sent about 100 phishing messages to 42 internal recipients. The agent contained it in four minutes and retrieved 92 outbound messages. Four higher-impact actions were held for human approval.
- Detection tiers: tier one under 500 ms for all mail, tier two for under 1% of messages, tier three for under 0.01%, with an organizational graph to correlate campaigns across mailboxes.
Remediation
- Keep signature, reputation and IOC-based detection and behavioral baselines, and add intent-based analysis of what a message is trying to get the recipient to do.
- Correlate similar requests across an organization and across mailboxes rather than scoring messages one at a time.
- Analyze text and images together to catch disruptive and harassment campaigns.
- Verify bank-detail changes and payment requests out of band, especially when the sender discourages verification or asks to move off the thread.
- Use automated investigation across email, identity, messaging and browser telemetry, with human approval for high-impact actions such as credential resets.