Google Threat Intelligence Group (GTIG), Google's threat intelligence unit, has published a report on how attackers are using AI, titled "GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold," on the official Google Cloud blog (dated November 17, 2026). GTIG says that in the past quarter it identified, for the first time, threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop. Mandiant also took part in the research.
exploit_agent Turns Patch Diffs Into Exploits and Runs Them
On an exposed staging server belonging to an attacker, GTIG found a multi-agent framework for vulnerability research and weaponization, which it calls exploit_agent. The framework integrates with multiple frontier and local LLMs.
Its workflow runs as follows:
- It compares the code of popular web applications before and after patch announcements
- It analyzes the differences to pinpoint potential vulnerabilities
- It drafts exploit code
- It automatically executes those exploits without human intervention
The server also held post-compromise scripts for internal reconnaissance, brute-forcing credentials and installing web shells, and even for removing web shells left by other attackers. According to GTIG, victim telemetry on the server suggests the operator was focused on French websites in the media and entertainment industries, as well as the education sector.
Open-Source Pentest Tools Are Being Repurposed for Real Intrusions
GTIG also cited cases in which open-source autonomous penetration-testing frameworks, built for defenders, were used in real attacks.
- CAIRN: A framework in which multiple agents share information as they work. GTIG observed it used in intrusions against targets including the US aerospace sector
- PentAGI: GTIG tracked an attacker using a customized attack platform built from the open-source version from late August to early October 2026. The targets were organizations in Asia
- ARTEX: Multiple threat actors have abused it in live intrusions
At the same time, GTIG notes that these tools trade stealth for speed. Letting them run autonomously without limits generates a large amount of network traffic that investigators can use as forensic evidence.
Zero-Day Exploits Spiked to 22 in August 2026
The report also draws on an analysis GTIG published in September 2026. Between January and August 2026, the number of vulnerability disclosures and exploits nearly doubled. Zero-day exploits rose only marginally, from a 2025 average of 8 per month to 11 per month in 2026, before spiking to 22 in August. The report cites these figures as data on vulnerability discovery and exploitation trends in the AI era.
Developer AI Tools and Other People's LLM Access Are Also Targets
Among North Korean threat activity, the report describes the tactics of a group suspected to be MIDNIGHT NEPTUNE. The actor created fake recruiter profiles on professional networking platforms, and the profile images contained SynthID watermarks, the digital watermark Google applies to AI-generated images. After gaining access, the actor modified .claude/settings.json configuration files in organizations' internal repositories so that when users started Anthropic's Claude Code, it would automatically trigger the download of the SOMBERMEME backdoor. GTIG says the actor also generated GitHub Personal Access Tokens named after Microsoft Copilot.
On "LLMJacking," in which attackers steal cloud credentials and run up charges on someone else's LLM access, Mandiant responded to several incidents in 2026. In some cases, single-day fraudulent charges exceeded $100,000, and in the most severe cases, cumulative abuse reached $1 million before remediation.
On influence operations, GTIG says that in August 2026 it investigated activity that combined the Gemini API with other AI services to build an end-to-end platform aimed at public sentiment in Europe and the Middle East. A separate contractor used the Gemini API to build an influence platform handling more than 100,000 accounts. Google says it disabled the assets used in the abuse and strengthened safeguards such as its classifiers.
Our Take: Attack Speed Is Eroding Defenders' Assumptions
Until now, AI misuse has mostly meant helping humans with tasks such as writing phishing messages or doing research. In our view, exploit_agent marks a step further, because a machine runs the whole process from writing exploit code to executing it, triggered by the release of a patch. The window between a fix and the first attacks is likely to shrink further, and organizations that roll out updates on a weekly or monthly cycle will probably be the most exposed.
The Claude Code settings-file technique also deserves attention. Configuration for AI coding tools often sits in the same repository as the code, which makes tampering easy to miss. We think reviewing changes to these settings files just as carefully as code changes will become essential. On the other hand, autonomous tools that leave plenty of traces are also easier targets for organizations that have solid monitoring in place.
Summary
In its latest report, GTIG says it has identified, for the first time, attackers using LLM frameworks to automate everything from vulnerability discovery to exploitation without human involvement. The report also covers the repurposing of open-source pentest tools, intrusions that abused Claude Code settings files, and LLMJacking cases with single-day charges of more than $100,000. Zero-day exploits reached 22 in August 2026, and Google says it has strengthened its defenses against misuse of Gemini.