Skip to navigation

Threat reportThreat IntelligenceTL-2026-2668

AI-Powered Cyber Attacks: Emerging TTPs Across Phishing, Deepfake BEC, Polymorphic Malware, and Prompt Injection

mediumACTIVE

Multiple corroborating 2023-2026 vendor and government reports document AI/LLM tooling now embedded across the full attack lifecycle: LLM-personalized phishing built from scraped LinkedIn/breach data, deepfake voice/video used to defraud Arup of $25.6M in a single Hong Kong incident, self-rewriting AI-generated malware (BlackMamba, MalTerminal, PromptLock, LAMEHUG/PROMPTSTEAL) that evades signature and even EDR detection, state-nexus actors (Midnight Blizzard-linked, Chinese exploit-foundry operators, APT31/APT42/UNC2970) abusing Claude and Gemini for reconnaissance and autonomous exploitation, and zero-click prompt injection (EchoLeak, CVE-2025-32711) that exfiltrated data from Microsoft …

Key facts

IdentifiersTL-2026-2668, CVE-2025-32711
Also known asEchoLeak, Vibe Hacking (AI-directed autonomous exploitation)
SeverityMEDIUM
EPSSCVE-2025-32711: 8.0% (94.5th percentile)
Affected productsMicrosoft 365 Copilot (Word, Excel, PowerPoint, Outlook, Teams integrations); Multiple Enterprise LLM assistants/agents that ingest untrusted email or documents
Weaknesses (CWE)CWE-74
Nation-state nexusRussia, China, Iran, North Korea
MalwareBlackMamba, DarkSword, GiftDrop, LAMEHUG, LAMEHUG - S9035, MalTerminal, MiniPlasma, PowerChrome, PromptLock, shadow
ToolsBeEF, PentAGI
MITRE ATT&CK / ATLAS techniquesAML.T0040 AI Model Inference API Access, AML.T0051 LLM Prompt Injection, AML.T0054 LLM Jailbreak, T1027 Obfuscated Files or Information, T1059 Command and Scripting Interpreter, T1071 Application Layer Protocol, T1114 Email Collection, T1195 Supply Chain Compromise, T1204 User Execution, T1552 Unsecured Credentials, T1566 Phishing, T1567 Exfiltration Over Web Service and 7 more in the attack flow
MotivationESPIONAGE
Target sectorsgovernment administration, defense, finance, technology, health, energy, retail, education, manufacturing
Target regionsNorth America, Europe, Asia-Pacific, Global
Detection rules9 (1 shown free)
Sources16 cited sources

Detection coverage

As of 2026-09-26, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-2668 across Splunk SPL, Microsoft KQL and Sigma, covering 27 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: AI-Personalized Phishing & Device-Code Sign-In Impersonation — SPL

Detects Azure AD OAuth device-code authentication flows correlated with inbound email containing high-pressure lure language or originating from known AI-driven phishing infrastructure (ms365-live.com / teams.ms365-live.com), consistent with LLM-personalized spear-phishing feeding device-code account takeover as documented for GTG-20006, UNC2970 (Gemini-assisted recruiter-persona phishing), and APT42.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1598, T1566, T1656, T1585, T1589, T1591, T1593
  • Data sources: azure_ad_signin_logs, microsoft_365_defender_email_events, o365_management_activity_api, dns_logs
  • Severity: high
  • Confidence: high

Splunk SPL

index=o365 sourcetype="o365:management:activity" Workload=AzureActiveDirectory
| search (AuthenticationProtocol="deviceCode" OR ClientAppUsed="Device Code" OR AppDisplayName="Microsoft Authentication Broker")
| eval auth_client=coalesce(ClientAppUsed,AppDisplayName)
| join type=left UserId
    [ search index=o365 sourcetype="o365:management:activity" Workload=Exchange Operation=MessageTrace
      | rex field=Subject "(?<lure_marker>(?i)urgent|invoice|payroll|salary update|recruiter)"
      | eval phish_domain=if(match(SenderAddress,"(?i)ms365-live\.com"),1,0)
      | table UserId Subject SenderAddress lure_marker phish_domain ]
| where isnotnull(lure_marker) OR phish_domain=1 OR match(auth_client,"(?i)ms365-live\.com|teams\.ms365-live\.com")
| stats count min(_time) as first_seen max(_time) as last_seen values(SenderAddress) as sender values(auth_client) as auth_client values(Subject) as subjects by UserId, src_ip
| where count > 0

Microsoft KQL

let PhishDomains = dynamic(["ms365-live.com","teams.ms365-live.com"]);
SigninLogs
| where TimeGenerated > ago(24h)
| where AuthenticationProtocol =~ "deviceCode" or ClientAppUsed has "Device Code"
| extend AppUsed = tostring(AppDisplayName)
| join kind=inner (
    EmailEvents
    | where TimeGenerated > ago(24h)
    | where SenderFromDomain has_any (PhishDomains) or Subject matches regex @"(?i)urgent|invoice|payroll|salary update|recruiter"
    | project SenderFromDomain, Subject, RecipientEmailAddress, NetworkMessageId
) on $left.UserPrincipalName == $right.RecipientEmailAddress
| project TimeGenerated, UserPrincipalName, AppUsed, SenderFromDomain, Subject, IPAddress, Location

Sigma

title: AI-Personalized Phishing Lure Correlated with OAuth Device-Code Sign-In
id: 7c1e5b3a-4f2d-4a6e-9c8b-1a2b3c4d5e01
status: experimental
description: Detects device-code OAuth authentication flows correlated with inbound email from known AI-generated phishing infrastructure or high-pressure lure subject lines, consistent with LLM-personalized spear-phishing and device-code account takeover (GTG-20006 pattern).
references:
    - https://cloud.google.com/blog/topics/threat-intelligence/threat-actor-usage-of-ai-tools
    - https://www.anthropic.com/threat-intelligence-report-september-2026
author: Threadlinqs Team
date: 2026-09-26
tags:
    - attack.reconnaissance
    - attack.t1598
    - attack.t1589
    - attack.t1591
    - attack.t1593
    - attack.initial-access
    - attack.t1566
    - attack.defense-evasion
    - attack.t1656
    - attack.resource-development
    - attack.t1585
    - attack.tool.wormgpt
    - attack.tool.fraudgpt
logsource:
    product: m365
    service: azuread
detection:
    selection_devicecode:
        AuthenticationProtocol: 'deviceCode'
    selection_lure:
        Subject|re: '(?i)urgent|invoice|payroll|salary update|recruiter'
    selection_domain:
        SenderFromDomain:
            - 'ms365-live.com'
            - 'teams.ms365-live.com'
    condition: selection_devicecode and (selection_lure or selection_domain)
falsepositives:
    - Legitimate device-code sign-in for headless/IoT devices combined with unrelated urgent email traffic
    - Help-desk or payroll bulk email campaigns using similar subject phrasing
level: high

Known false positives:

  • Legitimate device-code sign-in for headless/IoT/CLI tooling (Azure CLI, Teams Rooms) paired coincidentally with unrelated urgent-subject email traffic
  • Bulk payroll/HR email campaigns using similar high-pressure subject phrasing

Tuning notes: Scope the device-code selector to your tenant's known device-code client IDs to suppress expected automation, and refresh the lure-subject regex and phishing-domain list from ongoing HUNT-phase intake rather than hardcoding a static set.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • AI-Personalized Phishing & Device-Code Sign-In Impersonation — KQL · unlock with TL Creds
  • AI-Generated Polymorphic Malware Runtime Code Synthesis — KQL · unlock with TL Creds
  • Zero-Click Prompt Injection Exfiltration & Fraudulent AI-Reseller C2 — KQL · unlock with TL Creds
  • AI-Personalized Phishing & Device-Code Sign-In Impersonation — Sigma · unlock with TL Creds
  • AI-Generated Polymorphic Malware Runtime Code Synthesis — Sigma · unlock with TL Creds
  • Zero-Click Prompt Injection Exfiltration & Fraudulent AI-Reseller C2 — Sigma · unlock with TL Creds
  • AI-Generated Polymorphic Malware Runtime Code Synthesis — SPL · unlock with TL Creds
  • Zero-Click Prompt Injection Exfiltration & Fraudulent AI-Reseller C2 — SPL · unlock with TL Creds

Indicators of compromise

As of 2026-09-26, Threadlinqs Intelligence tracks 27 indicators of compromise for TL-2026-2668.

Sample indicators (3 of 27)

  • Domain: ms365-live[.]com
  • Domain: soraki[.]cc
  • Domain: policenationale[.]cc

Defanged, TLP:CLEAR. Full set of 27 indicators: unlock with TL Creds.

27 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Multiple corroborating 2023-2026 vendor and government reports document AI/LLM tooling now embedded across the full attack lifecycle: LLM-personalized phishing built from scraped LinkedIn/breach data, deepfake voice/video used to defraud Arup of $25.6M in a single Hong Kong incident, self-rewriting AI-generated malware (BlackMamba, MalTerminal, PromptLock, LAMEHUG/PROMPTSTEAL) that evades signature and even EDR detection, state-nexus actors (Midnight Blizzard-linked, Chinese exploit-foundry operators, APT31/APT42/UNC2970) abusing Claude and Gemini for reconnaissance and autonomous exploitation, and zero-click prompt injection (EchoLeak, CVE-2025-32711) that exfiltrated data from Microsoft 365 Copilot without any user interaction.

This threat aggregates six distinct, independently evidenced AI-enabled attack techniques that a September 25, 2026 SOC Prime overview ties together as a single detection-engineering trend, and which this research substantiates against primary vendor/government reporting rather than treating as a single incident.

1) AI-generated phishing and social engineering: threat actors now scrape LinkedIn profiles, breach dumps, and public code/cloud-metadata footprints with LLMs to auto-generate individualized lures at volume, collapsing what Trend Micro documented as a 'LinkedIn to tailored attack in 30 minutes' pipeline. Google's Threat Intelligence Group (GTIG) confirmed North Korea's UNC2970 used Gemini to research cybersecurity/defense-firm job roles and salary data to refine fake-recruiter phishing personas, while Iran's APT42 used Gemini for translation, persona development, and phishing-content drafting across multiple operation stages.

2) Deepfake voice/video BEC: in February 2024, engineering firm Arup's Hong Kong office was defrauded of HK$200 million (~US$25.6 million) across 15 wire transfers after a finance employee joined a video conference in which every other 'participant' -- impersonating the CFO and colleagues -- was a real-time deepfake, following an initial CFO-impersonation spear-phishing email. The fraud was only discovered when the employee separately contacted head office about the 'secret transaction.'

3) Polymorphic, AI-written malware: HYAS Labs' March 2023 BlackMamba proof-of-concept queries OpenAI's API at every execution to re-synthesize its keylogging payload via Python's exec(), producing a functionally identical but binarily unique sample each run and recording zero detections against a leading EDR in testing. SentinelLABS' LABScon25 research (September 2025) documented MalTerminal, the earliest known malware with an embedded LLM (GPT-4) call used to generate ransomware or reverse-shell code on demand, plus PromptLock ransomware (uses a locally hosted gpt-oss:20b model via the Ollama API to write malicious Lua on the fly) and APT28-linked LAMEHUG/PROMPTSTEAL (invokes Qwen2.5-Coder-32B-Instruct via the Hugging Face API to generate attacker commands from static prompts). A year-long SentinelLABS VirusTotal retrohunt found 7,000+ samples containing 6,000+ unique hardcoded LLM-provider API keys, which the researchers note is itself a detection opportunity.

4) AI-accelerated reconnaissance and employee profiling: beyond the Gemini cases above, Anthropic's September 2026 threat intelligence report documents a Chinese-speaking operation (GTG-10007, tracked to Changsha, Hunan) that used Claude as an 'agent swarm' orchestration layer to decompose reconnaissance across roughly 50 organizations into parallel subagents, and a Russian state-nexus operation (GTG-20006, linked to Midnight Blizzard) that used Claude for reconnaissance, phishing-infrastructure automation, and command execution against 20+ Ukrainian/European government, defense, and embassy targets -- notably having Claude autonomously rebuild its own malware components the moment security products flagged them.

5) AI-assisted, multi-turn BEC conversation: industry reporting (Proofpoint, Abnormal Security) describes BEC evolving from single impersonation emails into coordinated, multi-persona campaigns that build trust over several exchanges, with the AI adjusting tone and content based on victim replies -- consistent with Anthropic's report of Claude being used for real-time exploitation-pipeline decision-making rather than one-shot content generation.

6) Prompt injection against enterprise AI assistants: CVE-2025-32711 ('EchoLeak', CWE-74, disclosed by Aim Security and patched server-side by Microsoft in June 2025, CVSS 9.3 per Microsoft / 7.5 per NVD) is the first documented real-world zero-click prompt-injection exploit against a production LLM system: a single crafted email caused Microsoft 365 Copilot to access and exfiltrate internal file contents to an attacker-controlled endpoint with no user interaction, by chaining bypasses of the XPIA cross-prompt-injection classifier, Markdown link redaction, auto-fetched images, and an allow-listed Teams proxy. Anthropic's September 2026 report separately documents the inverse pattern -- an AI-supply-chain actor (GTG-50020) injecting malicious instructions into an AI vendor's own automated evaluation sandbox to steal production API keys, and a fraudulent-reseller operation (GTG-50021) harvesting Anthropic account credentials through look-alike discounted-access sites.

Taken together, these six vectors show AI is not a single new technique but a force multiplier applied across the existing ATT&CK kill chain -- reconnaissance, resource development, initial access, execution, defense evasion, credential access, collection, exfiltration, and C2 -- which is why this threat is scoped and published as a cross-cutting TTP-trend advisory rather than a single-actor incident.

MITRE ATT&CK / ATLAS techniques by tactic:

ML Model Access

AML.T0040 AI Model Inference API Access

execution

AML.T0051 LLM Prompt Injection

Defense Evasion

AML.T0054 LLM Jailbreak; T1027 Obfuscated Files or Information; T1684.001 Impersonation

Execution

T1059 Command and Scripting Interpreter; T1204 User Execution

Command and Control

T1071 Application Layer Protocol

Collection

T1114 Email Collection

Initial Access

T1195 Supply Chain Compromise; T1566 Phishing

Credential Access

T1552 Unsecured Credentials

Exfiltration

T1567 Exfiltration Over Web Service

Resource Development

T1585 Establish Accounts; T1588 Obtain Capabilities

Reconnaissance

T1589 Gather Victim Identity Information; T1591 Gather Victim Org Information; T1593 Search Open Websites/Domains; T1598 Phishing for Information

Remediation

Patches

  • Microsoft 365 Copilot server-side patch for CVE-2025-32711 (EchoLeak) -- applied automatically by Microsoft as of June 2025; no client-side action required, but tenants should confirm Purview/XPIA classifier logging is enabled.

Immediate actions

  • Enforce out-of-band callback verification using pre-established phone numbers (never numbers supplied within the suspect call/email) for any wire-transfer, payroll, or credential-reset request, even when 'approved' over a video or voice call.
  • Apply Microsoft's server-side fix for CVE-2025-32711 (EchoLeak) and audit any enterprise AI assistant that ingests untrusted email/documents for indirect prompt-injection exposure (auto-fetched remote images, markdown link rendering, proxy allow-lists).
  • Rotate and scope down AI-provider (Anthropic/OpenAI/Google) API keys and session tokens, especially any obtained through third-party resellers or unmanaged proxy services, per Anthropic's documented fraudulent-reseller credential-harvesting cases.

Workarounds

  • Restrict Copilot/enterprise-AI-assistant ability to auto-fetch remote images or render reference-style Markdown links from untrusted email/document content until XPIA-equivalent hardening is independently verified.
  • Require secondary human approval for any AI-agent-initiated financial transaction, credential change, or bulk data/mailbox export.

Longer-term hardening

  • Tune EDR/behavioral detection to flag LLM-generated code patterns (hardcoded provider API keys, embedded prompt strings, runtime code synthesis via exec()/eval()) rather than relying on static signatures, given confirmed AI-rewritten polymorphic payloads (BlackMamba, MalTerminal, PromptLock, LAMEHUG).
  • Adopt liveness-detection and multi-channel (not single video/voice channel) identity verification for executive-authorized financial or credential actions to counter deepfake BEC.
  • Establish AI/agentic-tool usage governance: scope what actions an AI agent may take autonomously, require human approval for financial/credential/bulk-export actions, and log all agent tool calls for audit.

Affected and fixed versions

  • Microsoft — 365 Copilot (Word, Excel, PowerPoint, Outlook, Teams integrations)
    Vulnerable versions: all tenants prior to the June 2025 server-side fix for CVE-2025-32711 (EchoLeak)
    Fixed in: server-side patched by Microsoft, June 2025 -- no client update required
  • Multiple — Enterprise LLM assistants/agents that ingest untrusted email or documents
    Vulnerable versions: any deployment lacking indirect-prompt-injection input sanitization or output-scope controls
    Fixed in: N/A -- requires architectural mitigation, not a vendor patch

Timeline

  • HYAS Labs discloses BlackMamba, a proof-of-concept polymorphic keylogger that calls OpenAI's API at every execution to re-synthesize its payload, recording zero detections in EDR testing.
  • WormGPT, a GPT-J-based uncensored LLM fine-tuned for BEC/phishing content generation, appears on underground forums marketed as 'the blackhat alternative to GPT.'
  • FraudGPT is identified circulating on dark web forums and Telegram channels for phishing pages, scam content, and vulnerability discovery.
  • Engineering firm Arup's Hong Kong office is defrauded of roughly HK$200 million (~US$25.6 million) across 15 wire transfers after a finance employee is deceived on a video call where every other participant, impersonating the CFO and colleagues, is a real-time deepfake.
  • Arup is publicly confirmed as the victim of the Hong Kong deepfake video-conference BEC scam.
  • CVE-2025-32711 ('EchoLeak') is disclosed by Aim Security and patched server-side by Microsoft: a zero-click prompt-injection flaw letting a single crafted email cause Microsoft 365 Copilot to exfiltrate internal data with no user interaction.
  • Academic paper formalizes EchoLeak as the first real-world zero-click prompt-injection exploit against a production LLM system.
  • SentinelLABS presents LABScon25 research documenting MalTerminal (earliest known LLM-embedded malware using GPT-4), PromptLock, and APT28-linked LAMEHUG/PROMPTSTEAL, plus a year-long VirusTotal retrohunt finding 7,000+ samples with 6,000+ hardcoded LLM API keys.
  • Google's Threat Intelligence Group discloses state-backed actors (China's APT31/Temp.HEX, Iran's APT42, North Korea's UNC2970, and Russian groups) abusing Gemini for reconnaissance, phishing-lure creation, exploit research, and C2 development across the full attack lifecycle.
  • A data leak reportedly exposes close to 19,000 WormGPT buyer accounts, including emails and payment metadata.
  • Anthropic publishes its September 2026 threat intelligence report detailing seven categories of AI misuse, including a Russian state-nexus operation (GTG-20006, Midnight Blizzard-linked) that had Claude autonomously rebuild detected malware, a Chinese exploit-foundry operation (GTG-10007) running agent-swarm reconnaissance across ~50 organizations, ShinyHunters-affiliated 'vibe hacking' (GTG-50014), and AI-supply-chain credential-theft campaigns (GTG-50020, GTG-50021).
  • SOC Prime publishes 'What Is an AI-Powered Cyber Attack?', synthesizing AI-personalized phishing, deepfake BEC, polymorphic malware, AI-accelerated recon, AI-assisted BEC dialogue, and prompt injection into a single detection-engineering trend brief.

Sources cited

Showing 15 of 16 sources, primary sources first.

Other reports on CVE-2025-32711

Previous and next in threat intel

Further reading

Drafted by the Threadlinqs research pipeline from 16 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats