# Google GTIG: AI Agents Exploit Patches, What to Check and Do Now

> Google's threat intelligence group says attackers now use AI agents to build exploits from patches. See what was reported, who is exposed and what to do.

Source: https://www.epifive.com/blog/google-gtig-ai-agent-exploits-patch-check · Published 2026-11-16 · EpiFive

## Key takeaways

- Google GTIG said on 16 November 2026 that it saw LLM frameworks fully automate vulnerability discovery and exploitation without a human in the loop for the first time.
- Google describes the patch-to-exploit window as shrinking, so treat a published patch as the start of the clock, not the end of the risk.
- Google says unauthorized enablement of cloud LLM services should be treated as a critical security event.
- The report publishes no standard indicator list, so do not wait for one before you review patches, LLM keys and cloud audit logs.

On 16 November 2026 the Google Threat Intelligence Group (GTIG) published its "AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold". It reports that threat actors are moving from manual prompting to open-source agentic offensive frameworks, and that GTIG saw LLM frameworks automate vulnerability discovery and exploitation end to end for the first time in the past quarter. There is no single patch to install. The most useful action is to shorten the time between a security update being released and installed on anything reachable from the internet, and to lock down the cloud and LLM credentials that attackers resell.

This guide separates what Google published from what others reported, copies the published identifiers exactly, and lists the defensive steps that Google and its documentation describe. It names no individuals and does not speculate about who is behind any activity.

## What happened

The primary source is the GTIG post on the Google Cloud Blog (Threat Intelligence), published 16 November 2026 by GTIG and Mandiant. The Hacker News reported on it the same day. Everything in the table below is Google's statement unless marked otherwise.

| Date | What the source says | Source |
|---|---|---|
| January to August 2026 | "while overall disclosures and exploits nearly doubled between January and August 2026, 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 the month of August." | GTIG post |
| June 2026 | The Jieyou phishing-as-a-service platform added an LLM chatbot that lets low-skill affiliates configure phishing panels in natural language. | GTIG post |
| August 2026 | An investigation found a Gemini-API-based platform for automating information-operations campaigns. | GTIG post |
| September 2026 | PaperCut NG/MF flaws CVE-2026-81578 and CVE-2026-82078 were exploited. GTIG says PaperCut was "reportedly" weaponized with LLM help within a day of disclosure; it also cites a GreyNoise report of a campaign four days after disclosure. | GTIG post |
| 21 October 2026 | PHP patches for CVE-2026-109618 and CVE-2026-109619 shipped. GTIG says patched code was public beforehand, which enabled code diffing, and that hundreds of likely LLM-generated, non-functional proof-of-concept exploits followed. | GTIG post |
| 16 November 2026 | GTIG publishes the AI Threat Tracker. | GTIG post |

### The headline finding

GTIG wrote: "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." It says it found an exposed staging server hosting a framework called `exploit_agent`, which performs continuous differential analysis of web application patches, drafts exploits and executes them autonomously. According to GTIG, victim telemetry suggested interest in French media, entertainment and education websites.

### Agentic frameworks

GTIG says it observed CAIRN and PentAGI in the wild and saw experimentation with CyberStrikeAI and HexStrike AI. It says ARTEX was also widely reported. Per GTIG, these frameworks share capabilities: discovering and exploiting low-complexity vulnerabilities, generating single-use payloads on demand and adjusting tactics from real-time feedback. Google's assessment: "Threat actor adoption of these frameworks represents a paradigm shift towards intrusions occurring at machine speed" and "they currently trade stealth for velocity."

### Other activity in the report

- **LLMjacking.** GTIG says Mandiant responded to several incidents in 2026: "In some cases, single-day fraudulent charges exceeded $100,000 USD. In the most severe cases, cumulative resource abuse reached $1 million USD before remediation." It says "LLM Mixers" resell stolen access at advertised prices of $0.10 to $25 per million tokens.
- **Information operations.** GTIG "assesses with high confidence that information operations (IO) threat actors have advanced from initial LLM experimentation to full operational integration."
- **State and criminal actors.** GTIG describes PRC-nexus, DPRK-linked and Iran-linked actors using LLM tools for reconnaissance, social engineering and fraud. This article does not repeat actor-level attribution beyond Google's own labels.

### Claims versus confirmed

| Statement | Who made it | Status |
|---|---|---|
| PaperCut flaws were weaponized with LLM help within a day | GTIG ("reportedly") | Google's own wording is hedged; treat as reported |
| Campaign four days after PaperCut disclosure, with foothold to domain admin "within minutes" | GreyNoise, as cited by GTIG | Third-party report cited by Google; we did not open the original |
| Fully automated exploitation with no human in the loop | GTIG | Google's assessment from an exposed server it found |
| Victim interest in French media, entertainment and education sites | GTIG | Based on "victim telemetry"; scale not stated |

## Am I affected?

Google does not give a victim count, a list of affected organizations or a single product to check. Whether you are exposed depends on three things.

1. **Internet-facing software with pending updates.** Google's finding is that attackers can diff patches, so any public-facing web application, print server or gateway that is behind on updates is the likeliest exposure. The report names these products and CVEs: PaperCut (CVE-2026-81578, CVE-2026-82078), PHP (CVE-2026-109618, CVE-2026-109619), the WordPress plugin Spectra (Ultimate Addons for Gutenberg) version 2.19.25 and LiteLLM (CVE-2026-40217, described as a sandbox escape that lets attackers elevate Virtual Keys to admin credentials).
2. **Cloud and LLM credentials.** If you use cloud LLM services, or a model gateway such as LiteLLM, stolen keys are the way LLMjacking starts. Google's LLMjacking post says that in multiple Mandiant incidents LLM services were not enabled at the time of compromise, and the attacker used stolen access keys to enable them.
3. **Developer tool configuration.** Google lists a backdoor called SOMBERMEME deployed through a modified `.claude/settings.json` file, which can auto-execute code.

### Identifiers Google published

Google's post does not publish IP addresses, domains or file hashes. The identifiers below are copied as listed; none is a live address or payload.

| Type | Identifier as published |
|---|---|
| Framework (found on an exposed server) | `exploit_agent` |
| Backdoor | SOMBERMEME |
| File path | `.claude/settings.json` (modified) |
| File name | `v21925.php` (Spectra plugin code Google discusses) |
| Behavior | GitHub Personal Access Tokens named after Microsoft Copilot |
| Behavior | "Burn and rotate": creating and deleting short-lived service accounts |
| Behavior | Automated aging of fake accounts before targeted messaging |

A filename or token name is a lead to investigate, not proof of compromise. Legitimate users also have `.claude/settings.json` files, so the signal is an unexpected change.

## What to do now

Actions are ordered by urgency. Items 1 and 2 follow from Google's findings; items 3 to 6 come from Google's LLMjacking guidance, which the AI Threat Tracker refers readers to.

1. **List internet-facing systems that have security updates waiting and install them first.** Start with the products named above if you run them.
2. **Treat a published patch as the start of the clock.** Google says patched code is public before some updates ship and that attackers use it. Do not wait for a monthly cycle for internet-facing software.
3. **Check whether cloud LLM services have been enabled that you did not enable.** Google says defenders should "treat any unauthorized enablement as a critical security event."
4. **Review and rotate cloud keys and tokens**, including cloud access keys, service account keys and GitHub personal access tokens, especially those you cannot account for.
5. **Set budgets, quotas and spending caps** on LLM services so abuse cannot run for hours unnoticed.
6. **Check developer machines and repositories** for unexpected changes to `.claude/settings.json` and other AI-tool configuration files.

### Step 3 in detail: audit logs Google says to watch

These operation names are copied from Google's LLMjacking post.

**AWS.** Monitor CloudTrail for events from `bedrock.amazonaws.com`. Execution: `Converse`, `InvokeModel` and `InvokeModelWithResponseStream`. Logging tampering: `GetModelInvocationLoggingConfiguration`, `PutModelInvocationLoggingConfiguration` and `DeleteModelInvocationLoggingConfiguration`. Google says "Any calls to disable logging by non-IaC (Infrastructure as Code) principals should be treated as high severity."

**Azure.** Monitor the `Microsoft.CognitiveServices` namespace. Google names `Microsoft.CognitiveServices/accounts/deployments/write` on the management plane for unauthorized new deployments, and notes that model invocations occur on the data plane and require Diagnostic Logs to be enabled.

**Google Cloud.** Google says inference operations are in Data Access logs, "which are disabled by default due to volume and ingestion costs", so you must configure audit policies to track them. It says to alert on enablement through `google.api.serviceusage.v1.ServiceUsage.EnableService`, and to watch `google.cloud.aiplatform.v1.PredictionService.Predict`, `GenerateContent` and `StreamGenerateContent`.

### Step 4 in detail: API keys

Google Cloud's API key best-practices page says to add restrictions ("By adding restrictions, you can limit the ways an API key can be used, reducing the impact of a compromised API key"), delete unneeded keys, rotate periodically ("Periodically create new API keys, update your applications to use the new API keys, and delete the old keys") and monitor usage ("Monitoring API usage can help alert you to unauthorized usage").

### If you find unexpected activity

Preserve the audit logs, revoke the affected credentials, and follow your provider's and your own incident process. This article does not provide provider-specific remediation commands because we did not open those pages in this run. Report US incidents through CISA's channels if appropriate for your organization.

## How to prevent it

Google's LLMjacking post says: "defenders should prioritize Identity and Access Management (IAM) and supply chain security". Its listed mitigations include the following, quoted or closely paraphrased.

- **Credential hygiene:** use a centralized secrets management platform to control management and rotation of keys and secrets.
- **Least privilege:** "Deny all permissions that are not explicitly required, including access to Cloud LLM services."
- **Code monitoring:** automate secret scanning in code repositories so API keys added to source are found and removed.
- **Budget alerts:** implement granular API quotas on LLM services and use spending caps.
- **MFA:** enforce phishing-resistant multi-factor authentication across cloud identity providers, consoles and developer platforms.
- **Granular API key scopes:** segment keys by model endpoint, environment and usage limits.
- **Gateway patch management:** extend vulnerability management to model proxy and routing gateways, such as LiteLLM, naming CVE-2026-40217.
- **Validation sweep monitoring:** watch for anomalous low-volume requests such as `max_tokens=1` or zero-temperature pings, which Google says attackers use for stealthy credential validation.
- **Sandboxed execution** for LLM applications that run generated code or database queries.

Google also says Mandiant Consulting released a hardening guide and architectural defense framework for LLM-enabled offensive tooling, and that Google SecOps deployed behavioral and volumetric rulepacks to detect ARTEX. We did not open the Mandiant guide.

## Checklist

- Internet-facing software with pending updates identified and scheduled.
- Products named in the report (PaperCut, PHP, Spectra, LiteLLM) checked if you run them.
- Cloud LLM services reviewed for enablement you did not approve.
- Audit logging for LLM inference turned on where it is off by default.
- Keys inventoried, restricted and rotated; secret scanning on.
- Budgets, quotas and spending caps set on LLM services.
- AI-tool configuration files in repositories reviewed for unexpected changes.

## What we checked

- Opened The Hacker News article on 2026-11-16 (IST) and confirmed it points to Google's post as the primary source.
- Opened the GTIG AI Threat Tracker on the Google Cloud Blog on 2026-11-16 (IST); quotes, dates and CVE numbers above are copied from it.
- Cross-checked the figures in the headline coverage with the primary post: the four named frameworks match, but the primary post says ARTEX was "widely reported" rather than observed by GTIG, so we used Google's wording.
- Opened Google's LLMjacking post and the Google Cloud API key best-practices page on 2026-11-16 (IST) for the mitigation and logging guidance.
- Confirmed that Google's post lists no IP addresses, domains or hashes; the table of identifiers contains only what it publishes.
- Claims from GreyNoise and PaperCut's weaponization timing are Google's citations; we did not open the original reports, and we did not open the Mandiant hardening guide or vendor advisories for the named CVEs.
- Not tested in any cloud account; the steps follow Google's published guidance as of the date below.

Last tested: 2026-11-16

## Sources

- Google Threat Intelligence Group, "AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold": https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
- Google Threat Intelligence Group, LLMjacking post: https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models
- Google Cloud, API key best practices: https://docs.cloud.google.com/docs/authentication/api-keys-best-practices
- The Hacker News, "Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches": https://thehackernews.com/2026/11/google-finds-hackers-using-ai-agents-to.html

## Frequently asked questions

### What did Google report about AI agents and exploits?

On 16 November 2026 the Google Threat Intelligence Group (GTIG) published its AI Threat Tracker. It says threat actors are using open-source agentic offensive frameworks, and that in the past quarter it identified actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.

### Is there a list of indicators of compromise I can search for?

Google's post does not publish a traditional list of IP addresses, domains or file hashes. It names tools, file paths and behaviors, which this article copies as published. Google's separate LLMjacking post lists cloud audit-log operations to monitor.

### What is LLMjacking and why does it matter to my business?

LLMjacking is the abuse of stolen cloud or LLM credentials to run models at the victim's expense. Google says Mandiant responded to several incidents in 2026 where single-day fraudulent charges exceeded $100,000 and, in the most severe cases, cumulative abuse reached $1 million before remediation.

### What is the most useful thing to do first?

Check which of your systems exposed to the internet have security updates waiting, and install them. Google says attackers can compare patched code with earlier versions, so the time between a patch and an attack is shrinking.

