Actively exploited Security news
Google GTIG: attackers fully automated vulnerability discovery and exploitation for the first time
On 16 November 2026 Google Threat Intelligence Group (GTIG) published its latest report on adversarial AI. For the first time, it says, it caught attackers using language models to fully automate both finding and exploiting vulnerabilities, with no human in the loop. The report also covers LLMJacking, where stolen access to AI services ran up bills of up to a million dollars. Here is what is confirmed and what it means for defenders.
Key points
- GTIG found a framework it calls exploit_agent that compares code before and after a patch, works out the flaw, drafts an exploit and runs it without human intervention. It turned up on an exposed staging server; the operator targeted French media, entertainment and education websites.
- Attackers are also running public offensive frameworks such as CAIRN, PentAGI, HexStrike AI and CyberStrikeAI. GTIG says intrusions now happen at machine speed but currently trade stealth for velocity, leaving clear traffic and code-level traces.
- Zero-day exploits rose from an average of 8 a month in 2025 to 11 a month in 2026, spiking to 22 in August 2026, according to GTIG.
- LLMJacking: Mandiant handled incidents where fraudulent AI charges topped $100,000 in a single day and, in the worst cases, reached $1 million before remediation.
What happened
Google Threat Intelligence Group (GTIG) published the latest edition of its AI Threat Tracker on 16 November 2026. The headline finding: in the past quarter GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop, for the first time.
The evidence is a framework GTIG calls exploit_agent, found on an attacker's exposed staging server. The operator went after French websites in media, entertainment and education. Its scripts handled internal reconnaissance, password brute-forcing, webshell installation and even removing webshells planted by rival attackers.
How it works
According to GTIG, exploit_agent runs a loop with no human step:
That shrinks the window between a fix going public and attacks starting. GTIG points to two cases from this year:
- PaperCut MF/NG (August 2026): a chain of two flaws (CVE-2026-81578 and CVE-2026-82078) allowed unauthenticated remote code execution. GTIG says researchers produced working exploits with LLMs within a day of disclosure, and four days later a campaign hit hundreds of organizations. NVD rates CVE-2026-81578 at CVSS 9.8; fixed versions are 24.1.10, 25.0.13 and 26.0.5.
- PHP (October 2026): patched code was public before the official fixes on 21 October 2026, letting attackers compare versions. Hundreds of non-working "proofs of concept", likely LLM-generated, followed and added verification work for defenders.
Alongside custom tooling, GTIG observed public offensive frameworks: CAIRN (used against targets including US aerospace), a customized PentAGI (reconnaissance and scanning of organizations in Asia from late August to early October 2026), plus HexStrike AI and CyberStrikeAI. A China-nexus group is building its own framework, likely for espionage.
LLMJacking: running up someone else's AI bill
The second major theme is theft of access to AI services. Attackers harvest keys leaked in public GitHub repositories or steal session tokens through deceptive websites. GTIG also describes abuse of CVE-2026-40217 in the LiteLLM proxy to turn limited "Virtual Keys" into full admin credentials. NVD describes the bug as code execution via /guardrails/test_custom_code in LiteLLM through the 2026-04-08 build (CVSS 8.8).
The damage is financial. Mandiant handled incidents with fraudulent charges above $100,000 in a single day, reaching $1 million in the worst cases. In one case a botnet fired off millions of API calls in roughly 24 hours before the keys were revoked. Stolen access is resold through "LLM mixers" on Russian- and Chinese-language markets for $0.10 to $25 per million tokens.
Other findings
- Iran's CALANQUE ION (formerly tracked as APT42) uses models including Gemini for reconnaissance and targeted social engineering.
- MIDNIGHT NEPTUNE, a suspected North Korea-linked cluster, targeted cryptocurrency firms in North America and Europe with fake recruiters using AI-generated photos and ClickFix lures. It also modified
.claude/settings.jsonso that starting Claude Code downloaded the SOMBERMEME backdoor. - GTIG assesses with high confidence that influence operations (Russia, China, Iran and commercial vendors) have moved from experimenting with LLMs to full operational use, including one platform managing more than 100,000 accounts.
How to spot compromise
GTIG did not publish hashes, domains or IP addresses in this report. It does note that today's AI frameworks favor speed over stealth: they generate noticeable traffic, rely on recognizable identifiers and leave code-level traces. Our suggestions:
- Look in web server logs for large volumes of varied requests from one source in a short window, especially right after a patch for a product you run.
- For AI services, watch for sudden spikes in usage and billing, short-lived service accounts that are created and deleted, and unexpected admin privileges.
- On developer machines, review AI assistant configuration files (such as
.claude/settings.json) and GitHub personal access tokens nobody remembers creating.
What to do
- Patch faster. Update internet-facing systems within days of a fix. Our CVSS calculator helps you rank what to patch first.
- Update LiteLLM to a build later than 8 April 2026 and review issued Virtual Keys.
- Keep AI keys out of code, monitor public repositories, and set budget caps and usage alerts.
- The basics still work: strong passwords with multi-factor authentication and closing unneeded services stop AI-driven attacks too.
We covered similar AI agent operations in our reports on Cairn and CL-STA-1238 and ARTEX and Storm-3358.
Timeline
Updates
- 16 Nov 2026 - first version.
FAQ
Is AI now attacking entirely on its own?
In one documented case. The no-human-in-the-loop claim rests on exploit_agent, found on a single exposed server run by an unidentified operator. Other frameworks are still steered by people, but AI is compressing reconnaissance and exploitation dramatically. GTIG expects operations to move further toward fully agentic intrusions.
What is LLMJacking?
Stealing access to paid AI models, through keys leaked on GitHub, hijacked sessions or flaws in proxies such as LiteLLM. Attackers then burn through the victim's quota or resell the access through so-called LLM mixers. Victims usually notice only when the bill arrives.
What should defenders change?
Mostly patch speed. GTIG says attackers use AI to diff patched and unpatched code and can have a working exploit within hours or days. For AI services, monitor usage, keep keys out of repositories and watch for new service accounts.
Sources
Spotted a mistake? Let us know and we will correct it openly.