Agentic Threat Groups: From Prompts to Autonomous Intrusions

Google Threat Intelligence Group (GTIG) reports attackers now run multi-agent frameworks like ARTEX, CAIRN, and PentAGI in the wild. Key trends and defences.

William OGOU
14 min read
Share
On this page9 sections

Adversarial artificial intelligence crossed a decisive threshold in late 2026: threat groups no longer rely on Large Language Models (LLMs) merely to write phishing lures or debug scripts. According to the November 16, 2026 Threat Tracker update from Google Threat Intelligence Group (GTIG) and Mandiant, financially motivated crews and state-backed espionage clusters now deploy autonomous multi-agent frameworks to execute live intrusions at machine speed.

The underlying mechanics of compromise remain familiar: stolen credentials, exposed management interfaces, and unpatched perimeter systems. What changed is velocity and automation. Open-source offensive harnesses, human-out-of-the-loop exploit pipelines, and stolen cloud AI quotas let small operator teams chain reconnaissance, privilege escalation, and lateral movement in minutes rather than days.

What to Remember

  • Multi-agent frameworks in the wild: Attackers now operationalise open-source agentic penetration-testing harnesses including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI during real intrusions.
  • Stage 5 of adversarial AI maturity: GTIG assesses that threat actors now execute multiple attack-lifecycle stages autonomously with minimal human oversight, projecting stealthier self-hosted operations within three to six months.
  • Zero-human n-day weaponisation: An exposed exploit_agent server automated patch diffing, exploit generation, and execution without human intervention, driving n-day exploitation surges while zero-day averages moved modestly from 8 per month in 2025 to 11 per month in 2026.
  • LLMJacking fuels illicit marketplaces: Compromised enterprise cloud accounts generated over $100,000 USD per day in fraudulent inference bills (reaching $1 million USD cumulative) to feed underground LLM Mixers such as APIMart and JetNight.
  • AI coding tools become lateral vectors: State-linked operators poisoned .claude/settings.json repository configurations to backdoor developer machines when Claude Code launched, while Phishing-as-a-Service (PhaaS) platforms added natural-language LLM assistants.

How Threat Groups Evolved Beyond Prompting

Within twelve months, adversarial AI adoption moved from ad hoc chatbot queries to orchestrated multi-agent systems executing sequential intrusion phases. In November 2025, an early semi-autonomous precedent emerged when GTG-1002 hijacked Claude Code to automate reconnaissance and lateral movement. By October 2026, off-the-shelf agentic harnesses removed much of the custom engineering burden.

Evolution of threat actor use of LLMs (Source: GTIG)

Figure: GTIG’s seven-stage adversarial LLM maturity curve. Threat activity currently sits at Stage 5, where LLM-driven tooling executes multiple intrusion phases in the wild with minimal human intervention, while Stages 6 and 7 project fully autonomous campaigns over the next one to six months.

Maturity StageOperational CapabilityHuman RoleObserved Timeframe
Stage 1: PromptingDrafting phishing lures, debugging malware code, light target researchHuman-driven chatbotLate 2025 to early 2026
Stage 2: Malware CallsRuntime LLM API calls embedded in families (HONESTCUE, PROMPTSPY) for dynamic obfuscationAutomated runtime callEarly 2026
Stage 3: Task AutomationAssisting exploit creation and environment setup, routed via early LLM proxiesLLM-in-the-loopEarly to September 2026
Stage 4: Custom HarnessesExperimental autonomous orchestration via model APIs and Telegram botsExperimental autonomyMid to September 2026
Stage 5: Wild Multi-StageOpen-source offensive frameworks executing multiple intrusion phases in live compromisesMinimal human-in-the-loopCurrent stage (Late 2026)
Stage 6: No-Human SwarmsUnattended, noisy multi-agent attacks running end-to-end chains from target selection to goalNo human-in-the-loopProjected: 1 to 3 months
Stage 7: Stealthy AutonomyFully autonomous intrusions using obscure infrastructure or self-hosted models to evade detectionFully autonomousProjected: 3 to 6 months

Why Current Autonomous Agents Still Trade Stealth for Speed

At Stage 5, autonomous offensive frameworks remain forensically loud. In Mandiant incident response cases detailed in GTIG’s companion ARTEX analysis, multi-agent execution produced burst patterns humans cannot replicate: 27 Amazon Web Services (AWS) permission probes in 23 seconds, iterative failed SQL-to-JSP webshell writes generating 240 log events, and incremental probe filenames (artex-probe-*).

LLMs also hallucinate targets or misread permissions. Consequently, stealth-focused espionage groups still apply autonomous frameworks selectively, whereas financially motivated operators embrace noisy velocity because campaign scale offsets lower per-target stealth. That balance will shift as attackers adopt unmonitored inference routers like ZeroHollow and self-hosted agents like Agent Zero.

Multi-Agent Offensive Frameworks Observed in the Wild

Rather than building custom automation from scratch, threat groups increasingly repurpose open-source offensive frameworks designed for penetration testing. An orchestration agent receives a high-level objective and coordinates specialised sub-agents across reconnaissance, initial access, persistence, privilege escalation, and exfiltration.

Multi-agent LLM offensive frameworks (Source: GTIG)

Figure: Core architecture of agentic offensive frameworks. A single orchestrator translates tactical goals into iterative feedback loops across five intrusion stages, generating single-use payloads on demand and adapting when exploits fail.

GTIG and Mandiant documented five primary multi-agent frameworks across active intrusions and actor experimentation between August and November 2026:

FrameworkArchitecture and MechanismConfirmed Threat Activity
ARTEXGo Command and Control (C2) backend with Next.js console coordinating six specialised sub-agents across commercial and open-weight LLMsAdopted by extortion crews and suspected PRC-nexus cluster UNC7228 (STORM-0940), moving from initial credential access to domain compromise in 15 minutes to 4 hours
CAIRNMulti-agent penetration-testing platform built on an Observe, Orient, Decide, Act (OODA) loop sharing state via a central blackboardObserved by GTIG in live intrusion activity targeting US organisations, including the aerospace sector
PentAGIContainerised offensive orchestrator running Kali Linux tools (nmap, metasploit, sqlmap) inside isolated sandboxes across 12+ LLM backendsCustomised fork observed between August and October 2026 conducting systematic scans and intrusions against Asian telecoms, integrators, and aerospace targets
CyberStrikeAIGo-based orchestration API linking LLMs to over 100 security tools, YAML attack recipes, and custom scripts from a single consoleActive threat actor experimentation documented by GTIG for end-to-end attack-chain automation
HexStrike AIModel Context Protocol (MCP) offensive harness exposing 150+ security utilities as callable functions for 12+ autonomous sub-agentsActive threat actor experimentation spanning reconnaissance, web testing, exploitation, and binary analysis

Zero-Human Exploit Pipelines and N-Day Compression

Autonomous vulnerability research has moved from laboratory benchmarks to unattended exploitation infrastructure. For attackers, pointing LLMs at recently patched flaws is far more economical than hunting unknown zero-days from scratch: comparing pre-patch and post-patch source code gives an agent a bounded problem with immediate feedback.

Vulnerabilities exploited per month by n-days and zero-days (Source: GTIG)

Figure: Monthly exploited vulnerabilities tracked by GTIG from January 2025 through August 2026. Overall exploitation nearly doubled in 2026, driven largely by rapid n-day weaponisation alongside an August spike to 22 exploited zero-days (compared against a 2025 average of 8 per month and a 2026 average of 11 per month).

As examined in our breakdown of GTIG’s 2026 AI-era CVE trends, three real-world cases show how agentic tooling shrinks patch windows:

  • Unattended exploit_agent staging server: GTIG uncovered a Russian-commented multi-agent pipeline wired to both frontier and local LLM APIs. The exploit_agent server continuously diffed web-application updates (including access-control logic in the Spectra WordPress plugin v2.19.25), drafted exploits, executed them without human approval, dropped webshells, and removed rival attackers’ shells while targeting French media, entertainment, and education sites.
  • PaperCut rapid chain (CVE-2026-81578 and CVE-2026-82078): Within 24 hours of August 2026 disclosure, security researchers used LLMs to build working Remote Code Execution (RCE) exploits chaining an authentication bypass with unsafe dynamic class loading. Four days post-disclosure, an automated campaign hit hundreds of organisations and escalated from foothold to Domain Administrator within minutes.
  • PHP2Shell noise flood (CVE-2026-109618 and CVE-2026-109619): Following the October 2026 PHP2Shell RCE disclosure, early patched code enabled fast diffing while hundreds of non-functional, LLM-generated Proof-of-Concept (PoC) scripts flooded public channels, forcing defenders to debunk broken claims while hunting for genuine exploitation.

LLMJacking and the Underground Model Economy

Scaling autonomous agents and synthetic media costs tokens, money, and compute. To bypass billing, rate limits, and provider geographic restrictions, cybercriminals steal cloud credentials and inference keys directly from victim infrastructure, a tactic known as LLMJacking.

LLMJacking API burst profile over an eight-hour window (Source: GTIG)

Figure: Telemetry from a Mandiant LLMJacking investigation. Within an eight-hour intrusion window, hijacked credentials sustained 180 to 260 successful inference requests per second while 429 Rate limit bursts spiked to roughly 1,100 requests per second.

Unlike slow cryptojacking, LLMJacking generates extreme financial damage within hours: Mandiant documented single-day fraudulent cloud bills exceeding $100,000 USD and cumulative losses reaching $1 million USD before remediation. Across financial, retail, transport, and technology victims, GTIG observed consistent patterns:

  1. Credential and gateway targeting: Attackers harvested keys from public GitHub repositories, tricked victims via JSON token-pasting lures, and exploited a sandbox escape in LiteLLM (CVE-2026-40217) to elevate restricted Virtual Keys into full PROXY_ADMIN administrative credentials (compounding earlier LiteLLM gateway authentication flaws).
  2. Quota burn and GPU provisioning: Intruders enabled managed LLM services on compromised cloud projects, rotated short-lived service accounts to dodge rate caps, and requested GPU quota increases to run agentic frameworks inside victim tenants.
  3. Resale through LLM Mixers: Illicit proxy storefronts such as APIMart and JetNight pool stolen enterprise keys and free-trial accounts behind open-source gateways (new-api, gcli2api). These LLM Mixers spread buyer traffic across multiple victim accounts at $0.10 to $25 USD per million tokens (or roughly $60 USD per month flat-rate plans), lowering detection odds while funding downstream abuse.

State Espionage, Agent Hook Poisoning, and Turnkey Crime

Beyond offensive pentesting frameworks, threat groups across espionage, influence operations, and cybercrime now weave LLMs into daily workflows and enterprise AI attack surfaces.

State-Sponsored Workflows and Dead-Drop Command and Control (C2)

Iranian group CALANQUE ION (APT42) continues using LLMs for target reconnaissance and tailored social engineering. Among People’s Republic of China (PRC)-nexus actors, contractor LEAKY CASTLE leveraged LLMs to translate and synthesise stolen documents into structured organisational intelligence reports, while other PRC clusters abused frontier LLM file-hosting backends as dead-drop C2 servers and intranet exfiltration pipes.

Supply Chain Poisoning via AI Coding Assistants

North Korean (DPRK) operators evolved beyond generating fake resumes and real-time screening answers for remote IT worker fraud. In compromises targeting cryptocurrency firms in North America and Europe, suspected cluster MIDNIGHT NEPTUNE created SynthID-watermarked fake recruiter profiles that directed victims to “ClickFix” lures.

Once inside internal code repositories, MIDNIGHT NEPTUNE weaponised developer AI tooling as a persistence mechanism. The operators modified .claude/settings.json inside repositories so that launching Claude Code automatically downloaded the SOMBERMEME backdoor, while naming rogue GitHub Personal Access Tokens (PATs) after Microsoft Copilot to blend in with legitimate automation.

Turnkey Influence Operations and Phishing Chatbots

In information operations, commercial Influence-as-a-Service (IaaS) contractors built multi-provider platforms automating more than 100,000 social accounts via cloud phone farms, residential proxies, automated account ageing, regional slang with deliberate typos, and smartphone-style photo flaws. Meanwhile, profit-driven political clickbait networks used LLMs to chase high-yield topics, and Chinese-language PhaaS platform Jieyou added a model-agnostic chatbot in June 2026 so low-skilled affiliates can customise phishing panels in plain language.

How Defenders Counter Machine-Speed Agentic Threats

Because autonomous offensive tools accelerate familiar techniques rather than inventing magical zero-days, effective defence rests on denying fast lateral paths and spotting volumetric anomalies early. Security engineering teams should prioritise five controls:

  1. Isolate control planes from the public internet: Place cloud consoles, CI/CD runners, hypervisors, backup managers, Kudu Source Control Manager (SCM) deployment endpoints, and AI gateways behind Zero Trust Network Access (ZTNA) and phishing-resistant Multi-Factor Authentication (MFA).
  2. Eliminate long-lived programmatic keys: Replace static cloud access keys, classic GitHub PATs, and shared storage keys with Workload Identity Federation (WIF), OpenID Connect (OIDC) federation, or short-lived managed identities bound to specific source networks and projects.
  3. Govern AI agent configurations and gateways: Enforce CODEOWNERS reviews and branch protection on repository agent configuration files (such as .claude/settings.json, .cursor, and Model Context Protocol (MCP) manifests covered in our IDE security analysis). Patch AI proxies like LiteLLM and Langflow promptly and isolate execution sandboxes from internal metadata endpoints.
  4. Tighten internet-facing patch timelines and suppress verbose errors: Expect public n-day disclosures to face automated diffing within days. Prioritise perimeter exposure over raw CVSS scores, and strip stack traces, SQL errors, and compiler output from external responses so autonomous agents cannot self-debug failed payloads.
  5. Alert on identity velocity and API billing anomalies: Instrument cloud audit logs for sudden Identity and Access Management (IAM) enumeration bursts, 429 Rate limit spikes, stealthy max_tokens=1 API key validation sweeps, unexpected EnableService calls for LLM endpoints, and outbound callbacks to disposable webhook collectors.
bash
# Quick audit: detect modified or unexpected Claude Code hook configs across local repositories
find . -maxdepth 4 -path '*/.claude/settings*.json' -exec grep -HnE 'hooks|command|curl|wget|bash|powershell' {} +

Conclusion: Defending Against Automated Velocity

The shift from manual prompting to Stage 5 multi-agent frameworks means defenders no longer compete solely against human typing speed. Open-source orchestrators like ARTEX, CAIRN, and PentAGI, unattended n-day pipelines like exploit_agent, and stolen quota marketplaces like LLM Mixers have industrialised reconnaissance, weaponisation, and cloud compute abuse into a repeatable commodity pipeline.

Yet autonomous tooling introduces clear weaknesses today: noisy request bursts, rigid retry loops, and predictable behavioural footprints. Organisations that eliminate static secrets, isolate control planes, protect agent execution hooks, and deploy volumetric identity monitoring will blunt the machine-speed advantage long before Stage 7 stealth arrives.

To further enhance your cloud security and implement Zero Trust, contact me on LinkedIn Profile or [email protected].

FAQ

Frequently Asked Questions (FAQ)

Which autonomous multi-agent frameworks are attackers using in the wild?

Google Threat Intelligence Group (GTIG) and Mandiant observed threat actors actively deploying ARTEX, CAIRN, and PentAGI during live intrusions, alongside active experimentation with CyberStrikeAI, HexStrike AI, and Agent Zero. Agentic offensive frameworks use an orchestrator agent to dispatch specialised sub-agents across reconnaissance, exploitation, privilege escalation, and lateral movement.

How has threat actor use of LLMs evolved from 2025 to late 2026?

GTIG maps adversarial LLM adoption across seven stages. Attackers progressed from chatbot-style prompting in late 2025 (Stage 1) and runtime LLM API calls in malware (Stage 2) to Stage 5 in late 2026, where multi-agent tooling executes multiple intrusion phases in the wild with minimal human oversight. GTIG projects fully unattended, stealthier Stage 6 and Stage 7 operations within one to six months.

Are AI agents mostly discovering new zero-days or exploiting n-days?

GTIG data from January 2025 through August 2026 shows that n-day weaponisation drives most of the surge in exploitation, whereas monthly zero-day averages rose modestly from 8 in 2025 to 11 in 2026 (with an August 2026 spike to 22). Pipelines such as exploit_agent automate pre-patch and post-patch differential analysis because analysing disclosed fixes is cheaper and faster than open-ended zero-day discovery.

What is LLMJacking and how do LLM Mixers monetise stolen AI access?

LLMJacking is the compromise of enterprise cloud credentials or AI gateway keys to hijack foundation-model quotas and GPU compute, generating fraudulent charges exceeding $100,000 USD per day and up to $1 million USD cumulatively. Underground LLM Mixers such as APIMart and JetNight pool stolen enterprise keys behind proxy gateways and resell model access at $0.10 to $25 USD per million tokens.

How did MIDNIGHT NEPTUNE abuse AI coding tools during intrusions?

After gaining initial access via SynthID-watermarked fake recruiter profiles and ClickFix lures, the suspected DPRK cluster MIDNIGHT NEPTUNE modified .claude/settings.json files inside internal code repositories. When developers launched Claude Code, the tampered hook automatically downloaded the SOMBERMEME backdoor, while attackers disguised GitHub Personal Access Tokens under Microsoft Copilot naming.

What controls best defend against autonomous offensive AI agents?

Defenders should isolate administrative control planes behind Zero Trust Network Access (ZTNA) and phishing-resistant Multi-Factor Authentication (MFA), replace long-lived API keys with short-lived Workload Identity Federation (WIF) tokens, enforce strict code reviews on AI agent config files such as .claude/settings.json, patch AI gateways such as LiteLLM promptly, and alert on sudden identity-velocity or API rate-limit bursts.

Resources