GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold

Google Threat Intelligence Group (GTIG) and Mandiant‘s latest AI Threat Tracker reports that adversarial use of generative AI has reached a critical inflection point. Threat actors are moving beyond manual prompting to autonomous multi-agent offensive frameworks, fully automated vulnerability exploitation pipelines, and monetization schemes such as LLMJacking. The findings are grounded in Mandiant incident response engagements, global threat actor tracking, and live platform defenses.

GTIG observed multiple open-source multi-agent frameworks being abused in live intrusions, including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI. CAIRN, which coordinates specialized agents via a shared blackboard, was used against a US aerospace target. A customized version of PentAGI conducted reconnaissance and vulnerability scanning against organizations in Asia between late August and early October 2026, and a PRC-nexus actor was seen developing its own automated attack framework. These frameworks share capabilities such as low-complexity vulnerability discovery, on-demand payload generation, and real-time tactical adjustment. The report notes that current autonomous operations trade stealth for velocity, generating forensic traffic and recognizable code-level traces, but reduced human involvement and increased stealth are expected as tools mature. Google SecOps has deployed rulepacks to detect ARTEX, and Mandiant Consulting published a hardening guide against LLM-enabled offensive tooling.

For the first time, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop. An exposed staging server hosted a framework called exploit_agent that integrates frontier and local LLMs, continuously performs differential analysis of popular web applications before and after patch announcements, drafts exploits, and executes them automatically. Victim telemetry suggested interest in French media, entertainment, and education websites. Public campaigns corroborated the trend: PaperCut vulnerabilities CVE-2026-81578 and CVE-2026-82078 were exploited as zero-days, with researchers generating working exploits within a day and Greynoise reporting an LLM-assisted campaign four days after disclosure. PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619 were discovered using LLM-enabled analysis, and hundreds of claimed but non-functional POCs appeared after the patch. Overall disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose from a 2025 monthly average of 8 to 11 in 2026, then spiked to 22 in August.

GTIG assesses with high confidence that information operations actors have moved to full operational integration of LLMs across state-sponsored groups, commercial Influence-as-a-Service vendors, and inauthentic political monetization networks. LLMs are used to produce synthetic media, localized personas, and automated distribution; improved content quality does not necessarily mean greater impact. GTIG investigated activity using the Gemini API to build an end-to-end IaaS platform targeting Europe and the Middle East, with prompts designed for astroturfing, localized dialects, intentional typos, and simulated smartphone camera flaws. Another commercial contractor was building a platform to manage over 100,000 accounts using virtual smartphones, proxy routing, and LLM-simulated user behavior to age accounts. IPM networks are mass-producing political clickbait using LLM-generated content and synthetic news anchors.

In incident response, Mandiant tracked LLMJacking intrusions in which financially motivated actors compromised cloud environments to consume LLM quotas and compute. Single-day fraudulent charges sometimes exceeded $100,000, and cumulative abuse reached $1 million before remediation in severe cases. Common TTPs include credential theft, session hijacking, creation of unauthorized IAM and service accounts, exploitation of CVE-2026-40217 (a LiteLLM sandbox escape), and burn-and-rotate service accounts to evade rate limits. Actors monetized access through LLM Mixers, illicit storefronts that route buyer traffic through stolen credentials and proxy infrastructure; observed pricing ranged from $0.10 to $25 per million tokens.

State-sponsored actors continue to integrate LLMs into espionage. PRC-nexus actors used LLMs for reconnaissance, post-compromise data analysis, and abuse of LLM file-hosting and backend APIs for staging, dead drop C2, and exfiltration. DPRK actors used LLMs to generate resumes and pass job screenings. Suspected MIDNIGHT NEPTUNE intrusions against cryptocurrency organizations combined fake recruiter profiles with SynthID watermarks, ClickFix lures, and modification of .claude/settings.json to drop the SOMBERMEME backdoor when users launched Claude Code. Phishing-as-a-Service platforms such as Jieyou added LLM chatbots that let low-skilled affiliates configure phishing panels with natural language.

Google’s defenses described in the report include proactive model hardening, real-time defenses against model distillation, account and project disruption, the autonomous Google AI Threat Defense architecture, and Gemini 4 Carbon with layered cyber-misuse safeguards stress-tested by internal red teams and external partners such as Gray Swan and Trajectory Labs. Google also cites Big Sleep, CodeMender, the Secure AI Framework, the Fairwind Program, and enforcement coordinated with YouTube and Trust & Safety.

GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold

View Original