Google Threat Intelligence Group (GTIG) said in its latest AI threat tracker, published Nov. 17, that it has identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time. The report was written with Mandiant and draws on frontline incident response work and live threat tracking.

An exposed server shows how the pipeline works

GTIG found a staging server hosting a multi-agent framework it calls exploit_agent. The tool compares code before and after patch announcements, drafts exploits and runs them automatically. Scripts on the server covered internal reconnaissance and the removal of other attackers' webshells. Victim data suggested the operator wanted French media and entertainment sites, along with the education sector.

Patched bugs are easier prey than zero-days, GTIG said, because attackers can work from a finite set of material such as code versions and release notes. GTIG's own count shows zero-day exploits rose from an average of 8 per month in 2025 to 11 per month in 2026, then reached 22 in August.

The wider timeline is short. Rapid7's Q3 2026 threat report found that exploitation occurred on average three days after public disclosure. CrowdStrike recorded its fastest exploitation after a public proof-of-concept at 35 minutes.

Defenders use the same approach. Big Sleep, an AI agent from Google DeepMind and Google Project Zero, searches for unknown security vulnerabilities in software. It found the SQLite flaw CVE-2025-6965, which Google described as known only to threat actors and at risk of being exploited.

Open source agent tools reach live intrusions

GTIG saw CAIRN and PentAGI deployed in the wild and tracked experiments with CyberStrikeAI and HexStrike AI. Between late August and early October 2026, one actor ran a customized platform built from open source PentAGI to scan organizations across Asia. CAIRN turned up in intrusion activity against aerospace targets in the US.

PentAGI is open source, self-hosted and built by VXControl. Its project site lists 33.4k GitHub stars and a v2.2.0 release on Oct. 5, 2026. The README says the tool is meant for authorized testing only.

The speed carries a price for attackers. GTIG notes that "they currently trade stealth for velocity." Unrestricted autonomous execution leaves forensic traffic and code-level traces behind. Google SecOps has deployed rulepacks to detect ARTEX, and Mandiant Consulting released a hardening guide for LLM-enabled offensive tooling.

Stolen AI access turns into a market

Mandiant responded to several LLMJacking incidents this year. Single-day fraudulent charges topped $100,000 in some cases, and cumulative abuse reached $1 million before remediation. The report says LLMJacking can generate "significant usage and billing increases within hours." In one case, a botnet or proxy service made millions of LLM API calls in about 24 hours on a single set of credentials before they were revoked.

Security firm Sysdig coined the term LLMjacking in May 2024 and estimated the potential cost to a victim at over $46,000 per day. By September 2024 it put the figure at over $100,000 per day with the newest models.

Resale runs through "LLM Mixers," illicit storefronts that route buyers through pools of stolen credentials and compute. On Russian- and Chinese-language marketplaces, GTIG saw prices from $0.10 to $25 per million tokens. One vendor, APIMart, applies a 5x increase once a prompt exceeds 100,000 tokens. Another, JetNight, sells monthly subscriptions.

Influence operations and espionage groups follow

GTIG assesses with high confidence that information operations actors have moved from LLM experimentation to full operational integration. In August 2026 it investigated an end-to-end platform built with the Gemini API and tools from other AI providers. The platform targeted public sentiment on regional political topics, mainly in Europe and the Middle East. A separate contractor designed a platform to manage over 100,000 accounts.

On the espionage side, a suspected MIDNIGHT NEPTUNE cluster targeted cryptocurrency organizations in North America and Europe. After initial compromise, the actor modified .claude/settings.json files in internal code repositories. When users started up Claude Code, it would automatically trigger the download of the SOMBERMEME backdoor.

Google said the activity triggered Gemini's safety responses. The company disabled the assets tied to it and passed the findings to Google DeepMind to strengthen its classifiers and models.

Five Events that Could Move Bitcoin - Crypto Week | HODL FM NEWS
Bitcoin starts the week near $64,000 after a drop to $62,000. See how Fed speeches, Nvidia earnings, Fed minutes, Anthropic's IPO and Japan data could move crypto.
hodl-post-image

Disclaimer: All materials on this site are for informational purposes only. None of the material should be interpreted as investment advice. Cryptocurrency investments involve significant risk of loss. Please note that, despite the nature of much of the material created and hosted on this website, HODL FM operates as a media and informational platform, not a provider of financial advisory services. The opinions of authors and other contributors are their own and should not be taken as financial advice. If you require advice, HODL FM strongly recommends contacting a qualified industry professional.