Google Threat Intelligence Group (GTIG) and Mandiant say adversarial use of AI has reached a turning point. In their latest AI Threat Tracker update, the researchers report that threat actors across cybercrime, espionage, and information operations (IO) are moving beyond simple prompting toward autonomous multi-agent frameworks, automated exploitation pipelines, and money-making schemes such as LLMJacking.
The findings draw on Mandiant incident response, global threat actor tracking and live platform defenses.
Threat Actors Hand Intrusions to Multi-Agent Frameworks
GTIG has observed several threat actors abusing open-source agentic pentesting frameworks during live intrusions. Beyond the widely publicized ARTEX, these include CAIRN, CyberStrikeAI and PentAGI, along with experiments involving HexStrike AI and custom harnesses. CAIRN runs an OODA (Observe, Orient, Decide, Act) loop where agents share a “blackboard,” and it was used against US aerospace targets.
CyberStrikeAI connects LLMs to more than 100 security tools through a Go-based C2 backend. HexStrike AI links LLM clients to over 150 tools coordinated by more than 12 sub-agents. Between late August and early October 2026, one actor used a customized PentAGI build to scan organizations across Asia. A PRC-nexus group was also building its own framework, and actors tested ZeroHollow and Agent Zero.
For now, these tools trade stealth for speed and leave forensic traces. GTIG expects that to change as the underlying models mature. Google SecOps has released rule packs to detect ARTEX.
AI Threat Tracker Finds Exploitation With No Human in the Loop
For the first time, GTIG saw threat actors fully automate vulnerability discovery and exploitation. An exposed staging server hosted “exploit_agent,” which compares web application code before and after patches, drafts exploits, and runs them automatically. A Russian comment in its configuration file, translated, instructed the operator to “enter the keys yourself — right here, not in the chat.” Victim data pointed to French media, entertainment and education targets.
The wider numbers reflect the shift. Disclosures and exploits nearly doubled between January and August 2026. Zero-days rose from a 2025 average of 8 per month to 11 in 2026, then jumped to 22 in August. In late August, attackers chained PaperCut flaws CVE-2026-81578 and CVE-2026-82078. Huntress and other researchers built working exploits with LLMs within a day, and Greynoise reported an LLM-assisted campaign four days after disclosure that reached domain administrator within minutes.
The PHP flaws CVE-2026-109618 and CVE-2026-109619, patched Oct. 21, 2026, drew hundreds of non-functional, likely AI-generated proofs of concept.
LLMJacking Turns Stolen Credentials Into Revenue
Mandiant handled several LLMJacking incidents in 2026. Single-day fraudulent charges topped $100,000, and in the worst cases cumulative abuse reached $1 million. Attackers pulled credentials from public GitHub repositories, created rogue IAM accounts, used a “burn and rotate” tactic, and exploited CVE-2026-40217, a LiteLLM sandbox escape. In one case, a botnet using leaked credentials made millions of API calls in about 24 hours.
Stolen access is resold through “LLM Mixers.” Prices range from $0.10 to $25 per million tokens. APIMart charges five times more once a prompt passes 100,000 tokens, while JetNight sells monthly subscriptions.
Influence and Espionage Threat Actors Go All In
GTIG assesses with high confidence that IO actors tied to Russia, the PRC and Iran, commercial Influence-as-a-Service vendors and inauthentic political monetization networks now routinely use LLMs. In August 2026, one actor used the Gemini API to build a platform targeting Europe and the Middle East. Its personas used local slang, deliberate typos, and fake smartphone-style photos. Another contractor designed a platform to run over 100,000 accounts through cloud phone farms. An IaaS provider tracked since 2019 used LLMs to improve its content.
On the espionage side, Iran’s CALANQUE ION (APT42) uses Gemini for reconnaissance, and PRC contractor LEAKY CASTLE analyzes stolen data with LLMs. A suspected DPRK cluster, MIDNIGHT NEPTUNE, created recruiter profiles carrying SynthID watermarks, then modified .claude/settings.json files to drop the SOMBERMEME backdoor. Separately, the Chinese phishing service Jieyou, active since March 2026, added an LLM chatbot in June 2026.
Google says it disabled the assets and accounts tied to this activity and strengthened its classifiers. It also cites Big Sleep, CodeMender, AI Threat Defense, the Secure AI Framework, and safeguards in Gemini 4 Carbon tested by Gray Swan and Trajectory Labs. Vetted defenders can get access to dual-use capabilities through the Fairwind Program.

