How to Stop AI-Driven Attacks: A Company Guide to Google's Report
Google Threat Intelligence Group says adversary use of AI has moved from asking chatbots for help to autonomous, multi-agent intrusion tools, including the first exploit framework it saw find and use vulnerabilities with no human in the loop. Most of the report is about speed. But every attack it describes still needs one of a handful of footholds on a person's machine or in a repo. This guide maps each attack to what stops it.
The short version
| Attack in the report | What it needs from you | What stops it |
|---|---|---|
| Backdoored coding-agent settings (MIDNIGHT NEPTUNE) | A developer starting Claude Code in a repo whose settings were changed | Review agent settings like code; endpoint guard on coding agents |
| Fake recruiters, fake meetings, "fix" commands | Someone pasting a command or installing a "meeting app" | Paste and command guards; fake-meeting checks |
| LLMJacking: stolen AI and cloud keys | A key on public GitHub, or typed into a fake page | Secret scanning; key paste warnings; spending limits |
| Stealing data out through AI platforms | Malware on a machine, talking to an AI API | Flag AI-API traffic from programs that shouldn't make it |
| Autonomous exploitation of new bugs | An unpatched internet-facing system | Patch speed and attack-surface reduction (not a browser problem) |
1. Coding agents turned into installers
A North Korea-linked group, MIDNIGHT NEPTUNE, edited .claude/settings.json in company repositories so that starting Claude Code downloaded the SOMBERMEME backdoor. It also created GitHub tokens named after Microsoft Copilot to blend in.
- Review: put
.claude/,.mcp.json,.cursor/,.vscode/,.devcontainer/andCLAUDE.md/AGENTS.mdunder CODEOWNERS. A hook is code that runs. - Audit tokens: list your organisation's personal access tokens and look for ones named after tools (Copilot, CI, deploy) that nobody remembers creating.
- Scan: the free AI repo security scanner checks a repo or an organisation's public repos for these settings.
- Endpoint: PhishClean Desktop 0.3.1 stops a coding agent that starts a hidden download command. Full detail in AI coding agent backdoors.
2. Fake recruiters and fake meetings
The same group used recruiter profiles with AI-generated photos to lead targets to fake meeting platforms with ClickFix lures: "your camera isn't working, paste this into Terminal or the Run box". DPRK IT-worker schemes used AI-written resumes and interview help. None of this is new; AI makes it cheaper and more convincing.
- No real interview needs you to paste a command or install a meeting app from a link. Zoom, Teams and Meet work in the browser.
- Run take-home tests in a throwaway VM or a cloud dev environment, never on a work laptop.
- Where PhishClean helps: the extension warns, for free, when an interview page tells you to paste a "camera fix" command, and flags fake meeting-app downloads and fake coding-test repos in the trial and on Pro. Desktop stops the pasted PowerShell or mshta command itself, on every plan. See fake job interview malware.
3. LLMJacking: stolen keys, resold compute
Attackers harvest valid keys from public GitHub repos and steal session tokens through fake websites, then resell access through "LLM mixers". The report cites single-day fraudulent charges above $100,000 and $1 million in total before one victim caught it. A LiteLLM flaw (CVE-2026-40217) let limited "virtual keys" become admin credentials.
- Turn on GitHub secret scanning and push protection, and scan your public footprint the way an attacker does: AI repo security scanner.
- Set hard spending limits and usage alerts on every AI provider account. Rotate any key that was ever committed, even if the file is gone: it stays in git history.
- Patch LiteLLM and any other AI gateway you run, and don't expose its admin interface.
- Where PhishClean helps: from extension 1.13.0, Paste Guard (trial and Pro) asks before you paste an Anthropic, OpenAI, AWS, GitHub or other API key into a page that looks wrong or a "verify your key" page that borrows the provider's name. It stays quiet on the provider's own site and on ordinary tools that ask for a key. Desktop stops programs run from Downloads or Temp from opening the files where Claude Code, Codex, Gemini CLI, AWS and GitHub keep their credentials.
4. AI platforms as a getaway car
China-nexus actors used a frontier model provider's file hosting to stage lures and exfiltrate stolen data, a backend API to send out intranet documents, and LLM platforms as "dead drop" command and control. To a firewall, this looks like developers using AI.
- Know which teams and apps are supposed to call AI APIs. Route sanctioned use through a gateway with your own keys, so unsanctioned traffic stands out.
- Where PhishClean helps: Desktop 0.3.1 (with its service on) warns when a program running from Downloads, Temp, the Desktop, or a Windows script host such as mshta or wscript, contacts an AI provider's API. Editors, Python, Node and installed apps are left alone, so developers aren't flooded with warnings.
5. Exploits in days, not months
The report's sharpest finding: an exploit framework that diffs patches, writes exploits and runs them with no human involved, and a PaperCut campaign that hit hundreds of organisations four days after disclosure, going from foothold to domain admin "within minutes". Zero-days averaged 11 a month in 2026, then spiked to 22 in August.
No browser extension or desktop app fixes this. What does: patch internet-facing systems within days, not a monthly cycle; take admin interfaces off the internet; watch for new admin accounts and service accounts. The good news in the report is that these tools "trade stealth for velocity": they are noisy, and leave distinct traces for whoever is looking.
A one-page checklist
- CODEOWNERS on agent settings and instruction files; a CI check that flags new hooks and MCP servers.
- Secret scanning and push protection on; spending limits on every AI and cloud account; old keys rotated.
- Coding agents' permission prompts on; project MCP servers not auto-approved.
- Interviews and take-home tests in sandboxes; nobody pastes commands from a web page.
- An inventory of who calls AI APIs and from where.
- Internet-facing patches in days; admin panels off the internet.
- An endpoint guard on developer laptops, where agent backdoors and stolen-key malware actually run.
Sources
- Google Threat Intelligence Group, November 2026: every attack described above, the figures quoted and the PaperCut timeline (citing GreyNoise).
PhishClean for Teams
The browser extension and Desktop on every laptop, one dashboard, a 30-day pilot. Built for the attacks in this report.
See PhishClean for TeamsLast updated: