Report: AI Is Scaling Adversary Operations and Compressing Attack Timelines
Summary
A review of 2025–2026 reporting finds AI is helping threat actors scale familiar tactics, from reconnaissance and phishing to malware development and agentic operations, while pressuring defenders to respond faster.
Key points
- The report says AI is accelerating existing attack methods rather than fundamentally changing how most breaches begin, which still commonly involve phishing, credential abuse, and unpatched systems.
- CrowdStrike data cited in the report puts average attacker breakout time in 2025 at 29 minutes, with a fastest recorded case of 27 seconds.
- Reported uses include AI-assisted target research, multilingual phishing, fabricated personas, malware development, and code obfuscation.
- The report describes PROMPTSTEAL/LAMEHUG querying an LLM during execution and QUIETVAULT checking for local AI command-line tools to search for secrets.
- Anthropic reported an AI-orchestrated espionage campaign targeting roughly 30 organizations; the degree of autonomy claimed has been questioned by some independent researchers.
- The report warns that faster vulnerability disclosure and growing AI supply-chain dependencies add pressure and attack surface for organizations.
- It recommends faster operationalization of threat intelligence, prioritization by attacker techniques, and visibility into emerging signals.
Article Details
- Publisher
- Eclectic IQ
- Report Period
- 2025–2026
- Scope
- AI-enabled cyber operations, including attack speed, activity across the attack lifecycle, fraudulent IT-worker schemes, and AI-orchestrated operations.
- Key Statistics
- CrowdStrike reported an average attacker breakout time of 29 minutes in 2025; the fastest recorded case was 27 seconds.
- Operations attributed to AI-enabled adversaries rose 89% year-on-year, according to CrowdStrike.
- The median time for an initial-access actor to hand a foothold to a follow-on group was 22 seconds.
- UN reporting estimated that the fraudulent IT-worker scheme generated $250 million to $600 million per year across about 40 countries.
- Recommendations
- Shorten the path from new intelligence indicators to deployed detections and response tooling.
- Prioritize AI-enabled techniques and behaviors that reduce time to impact rather than relying primarily on attribution.
- Incorporate detection for runtime-LLM malware, agentic tooling, and synthetic-media lures.
Threat Actors
APT28Russian actor reported to have deployed PROMPTSTEAL against Ukraine.APT31Chinese actor reported to have used Gemini for vulnerability analysis against Western targets.APT41Chinese actor reported to have used Gemini for C++ and Golang development, support for the OSSTUN command-and-control framework, and code obfuscation.APT42Iranian actor reported to have used Gemini to research targets and build personas.GTG-1002Tracked actor cluster that Anthropic assessed with high confidence to be Chinese state-sponsored; Anthropic reported it attempted intrusions against roughly 30 global targets using Claude Code.Jasper SleetMicrosoft-tracked North Korean IT-worker operation reported to use generative AI to build identities and face-swapping tools to alter images in stolen identity documents.MuddyWaterIranian actor reported to have posed as a student to obtain Gemini's assistance in developing custom malware.UNC2970North Korean actor reported to have used Gemini to profile high-value targets.
Malware
LAMEHUGGTIG reported that the Russian actor APT28 deployed PROMPTSTEAL (also tracked as LAMEHUG) against Ukraine — its first observation of malware querying an LLM in live operations [9].PROMPTSTEALGTIG reported that the Russian actor APT28 deployed PROMPTSTEAL (also tracked as LAMEHUG) against Ukraine — its first observation of malware querying an LLM in live operations [9].QUIETVAULTMandiant’s M-Trends 2026 confirmed the pattern, noting the QUIETVAULT credential stealer checked infected systems for local AI command-line tools to run prompts that hunt for further secrets [4].
Vendors
AnthropicAcross reporting from Google [1], Microsoft [2], CrowdStrike [3], Mandiant [4], Anthropic [5] and OpenAI [6], a consistent picture has emerged: artificial intelligence is not fundamentally changing what adversaries canCrowdStrikeAcross reporting from Google [1], Microsoft [2], CrowdStrike [3], Mandiant [4], Anthropic [5] and OpenAI [6], a consistent picture has emerged: artificial intelligence is not fundamentally changing what adversaries canGoogleAcross reporting from Google [1], Microsoft [2], CrowdStrike [3], Mandiant [4], Anthropic [5] and OpenAI [6], a consistent picture has emerged: artificial intelligence is not fundamentally changing what adversaries canMandiantAcross reporting from Google [1], Microsoft [2], CrowdStrike [3], Mandiant [4], Anthropic [5] and OpenAI [6], a consistent picture has emerged: artificial intelligence is not fundamentally changing what adversaries canMicrosoftAcross reporting from Google [1], Microsoft [2], CrowdStrike [3], Mandiant [4], Anthropic [5] and OpenAI [6], a consistent picture has emerged: artificial intelligence is not fundamentally changing what adversaries canOpenAIAcross reporting from Google [1], Microsoft [2], CrowdStrike [3], Mandiant [4], Anthropic [5] and OpenAI [6], a consistent picture has emerged: artificial intelligence is not fundamentally changing what adversaries can
Products
ChatGPTOpenAI reported a China-aligned cluster using ChatGPT to generate phishing content in English, Chinese and Japanese [6].Claude CodeAccording to Anthropic, the actor manipulated its Claude Code tool into attempting intrusions against roughly thirty global targets, with the AI performing the majority of the work and humans intervening at only aGeminiGoogle's Threat Intelligence Group reported several state-backed groups using Gemini to accelerate this phase [4].
Tools
Countries
ChinaOpenAI reported a China-aligned cluster using ChatGPT to generate phishing content in English, Chinese and Japanese [6].IranNorth KoreaUkraineGTIG reported that the Russian actor APT28 deployed PROMPTSTEAL (also tracked as LAMEHUG) against Ukraine — its first observation of malware querying an LLM in live operations [9].