Inter-office memo

To
Everyone who runs a website or uses Claude Code
From
Mo, Chief Executive
Date
November 18, 2026, 12:13 AM CST
Re
Google’s AI Threat Tracker

Google found AI agents hacking websites on their own, and a Claude Code backdoor

On November 16, 2026, Google’s threat team said that for the first time it had seen attackers fully automate finding and exploiting web flaws, with a framework it calls exploit_agent. The same report says suspected North Korean hackers edited .claude/settings.json in companies’ repositories so Claude Code fetched a backdoor at startup. Patch fast, check those files, and guard your AI keys.

What Google found

On November 16, 2026, Google Threat Intelligence Group (GTIG) and Mandiant published their quarterly AI Threat Tracker, titled “Agentic Threat Activity Crosses a Significant Threshold.” Its central finding is one sentence: “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.”

That is a change from Google’s last report. On September 8, GTIG wrote that it “has not yet observed threat actors deploying fully autonomous pipelines against targets in the wild.”

Four findings in the new report matter to people who run websites or use coding agents:

  1. exploit_agent, a framework that turns software patches into attacks with no one at the keyboard.
  2. A Claude Code backdoor. Suspected North Korean hackers edited Claude Code’s settings in companies’ own repositories, so starting it downloaded malware.
  3. Stolen AI keys, resold. Attackers run up cloud and AI bills on other people’s accounts, in some cases over $100,000 in a day, and sell the access on.
  4. More AI hacking frameworks in real attacks, beyond the ARTEX framework the FBI warned about on November 13.

BleepingComputer adds that the researchers expect unsophisticated attacks needing “little to no human interaction” within three months, and “sophisticated, fully autonomous” ones within six. Those projections aren’t in the report’s text.

How exploit_agent works

GTIG found it on “an exposed staging server,” the attacker’s own. It is a “multi-agent vulnerability research and weaponization framework” that “integrates with multiple frontier and local LLMs,” and the report describes its loop: it “compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention.”

The same server held data stolen from victims and scripts for “internal reconnaissance, brute-forcing, installing webshells, and removing other attackers’ webshells.” Its victim records suggest the operator was after “French websites associated with the media and entertainment industries, as well as the education sector.” A configuration file, commented in Russian, reads in Google’s translation: “LLM API keys for auto-exploitation (exploit_agent pilot).”

The report’s one example shows the framework reading the access checks in version 2.19.25 of Spectra, a WordPress plugin now called Spectra Legacy, with more than a million active installs. Google doesn’t say the framework found or used a flaw in it. Eight Spectra releases since 2.19.25 each “addressed a security bug,” most recently 2.20.6 on November 3.

Working from patches is the point. Comparing two versions of the same code is a much smaller search than hunting for unknown flaws, Google says, and AI is “shrinking the patch window for organizations.” Its other examples show how small that window is: when attackers exploited two PaperCut flaws in late August, researchers using AI had working exploits “within a day,” and GreyNoise saw a campaign against “hundreds of organizations” four days after disclosure.

The Claude Code backdoor

The second finding is about the tools developers already use. GTIG says a suspected North Korean cluster it calls MIDNIGHT NEPTUNE went after cryptocurrency companies in North America and Europe in three steps:

  1. Fake recruiters. Profiles on professional networks whose photos carried SynthID watermarks, Google’s marker for AI-generated images.
  2. Fake meetings. The recruiters sent targets to meeting pages with “ClickFix” lures, which told them to run “diagnostic fix scripts.”
  3. The repositories. Once inside, the attackers went into internal code repositories, “modifying .claude/settings.json configuration files so that when users started up Claude Code, it would automatically trigger the download of the SOMBERMEME backdoor.” They also created GitHub personal access tokens “named after Microsoft Copilot” to blend in.

The report describes no flaw in Claude Code. The attackers were already inside, and used a file that teams commit and share. Google doesn’t say which setting they used, but Claude Code’s settings files can run commands through hooks, and Anthropic’s documentation warns that “command hooks execute shell commands with your full user permissions.”

Claude Code does check before running them. It “holds back hooks from every settings file” until you accept the trust dialog for a folder, but that trust is kept per repository, so in a repository you already trust, such as your company’s own, a hook someone adds later runs at your next start. The documentation describes no second prompt for it. And in claude -p and Agent SDK runs, Claude Code “never shows the dialog and treats the folder as trusted.”

Stolen AI keys, resold

Google calls the third trend LLMJacking: stealing AI credentials and cloud accounts, then using or reselling the access. “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.” In one case a botnet “generated millions of LLM API calls in approximately 24 hours before the credentials were revoked.”

The ways in are ordinary: “harvesting valid credentials from public GitHub repositories,” fake sites that trick people into handing over “active session tokens,” and tokens taken “directly from victim endpoints and continuous integration environments.” One attacker used CVE-2026-40217, which Google describes as “a sandbox escape in LiteLLM that allows an attacker to elevate limited-access inference keys (‘Virtual Keys’) into full administrative credentials.” GitLab’s advisory, published May 11, says it affects LiteLLM 1.81.8 up to 1.83.10, which fixed it.

The access ends up in “LLM Mixers,” underground storefronts selling AI model access at “$0.10 USD to $25 USD” per million tokens.

The frameworks Google names

Framework What it is What Google saw
exploit_agent Turns web software patches into exploits and runs them Aimed at French media, entertainment and education sites
ARTEX Open-source AI penetration-testing framework Already “highly-publicized”; see our memo
CAIRN Open-source; agents share a “blackboard” and work toward a goal Used in the wild, including against US aerospace
PentAGI Open-source autonomous attack framework A customized version scanned organizations in Asia, late August to early October
CyberStrikeAI Open-source; connects AI models to over 100 security tools Experimentation
HexStrike AI Open-source MCP harness with over 150 tools Experimentation

Google’s one consolation: these tools “currently trade stealth for velocity,” leaving traffic and traces that make them easier to catch.

What to do

If you run a website.

  1. Update quickly. exploit_agent starts from the patch, so the time between a fix being published and your site installing it is the opening. Turn on automatic updates for plugins and themes where you can.
  2. On WordPress with Spectra Legacy, update to 2.20.6.
  3. Look for what its scripts leave behind: files you didn’t upload, such as webshells, sign-ins you don’t recognize, and bursts of failed ones.

If you or your team use Claude Code.

  1. Read .claude/settings.json before your next session in each shared repository, and .claude/settings.local.json if it’s in git. Look for hooks that download or run anything you don’t recognize. git log -p -- .claude/ shows who changed them, and when.
  2. Run /hooks in Claude Code. It lists every hook with the file it came from.
  3. Start without hooks while you check: claude --settings '{"disableAllHooks": true}', which Anthropic says “takes precedence over project and local settings.”
  4. In CI and scripts that run claude -p, start with --bare or that same flag, as Anthropic’s documentation advises for repositories you didn’t write.
  5. Make changes to .claude/ need a review. An admin can go further with allowManagedHooksOnly in managed settings, which blocks every hook but the organization’s own.
  6. Don’t run a “fix” script a meeting page or recruiter gives you, and check your GitHub personal access tokens for any you didn’t make, including ones named for Copilot.

If you hold AI keys. Google’s LLMJacking guidance is to restrict keys “by specific model endpoints, environments (dev/prod), and usage limits,” set quotas and “spending caps,” automate secret scanning in your repositories, and require phishing-resistant multi-factor authentication. If you run LiteLLM, be on 1.83.10 or later, and not on 1.107.0, the poisoned October release. Keeping keys out of chats and files helps too: in Enjoy’s desktop app, which runs Claude Code and Codex on the plans you already have, project secrets let an agent use a key by name without the value appearing in the conversation.

What Google hasn’t said

  • Which AI models exploit_agent used, or whether any lab’s safeguards ever stopped it.
  • How many sites it hit, or how long it took from a patch to an attack.
  • Who runs it, beyond its Russian-language comments.
  • Which Claude Code setting the North Korean cluster changed, or how many companies it reached. We found no statement from Anthropic about the report as of November 18.
  • Where the three- and six-month projections come from. BleepingComputer attributes them to the researchers; the report’s text doesn’t include them.

We’ll update this memo if Google, Anthropic or the targeted sites add to it.

8.Questions.

What is exploit_agent?

It’s an attack framework Google found on a hacker’s exposed staging server. Several AI agents, running on “multiple frontier and local LLMs,” compare popular web software before and after each patch, draft exploits for what changed, and run them “without human intervention.” Data on the server pointed to French media, entertainment and education websites.

Is this really the first fully autonomous AI cyberattack?

It’s the first Google has seen. Its report says it found attackers fully automating discovery and exploitation “without a human in the loop for the first time,” and in September it had written that it had “not yet observed” such pipelines in the wild. Google hasn’t said how many sites were hit.

Was Claude Code hacked?

Not by Google’s account, which describes no flaw in Claude Code. Suspected North Korean hackers first got into crypto companies through fake recruiters and fake meeting pages, then edited .claude/settings.json in internal repositories so that starting Claude Code downloaded their SOMBERMEME backdoor.

How do I check my repositories’ Claude Code settings?

Read .claude/settings.json in each shared repository before your next session there, look for hooks that download or run anything you don’t recognize, and check who changed it in git. Inside Claude Code, /hooks lists every hook with the file it came from.

Which AI models did exploit_agent use?

Google doesn’t say. Its report says only that the framework “integrates with multiple frontier and local LLMs,” and shows Russian comments from a configuration file set up for the operator’s AI API keys.

How do I protect my website from AI hacking agents?

Update quickly. exploit_agent starts from the patch itself, so the time between a fix being published and your site installing it is the opening. Turn on automatic updates for plugins where you can, and look for admin accounts and files you didn’t create.

Enjoy Agentic Industries · Checked against its sources on