If your patch SLA for internet-facing software is still measured in weeks, Google's newest threat report says that window now belongs to the attacker. On November 16, Google Threat Intelligence Group reported that it had identified threat actors using LLM frameworks "to fully automate vulnerability discovery and exploitation without a human in the loop for the first time". The same report describes a North Korean cluster planting a backdoor through a repository's .claude/settings.json, and Mandiant incident data with single-day LLMJacking bills above $100,000.
Three things should change on your side. Shorten the patch clock for internet-facing n-days. Treat AI agent configuration files in your repositories as executable code that needs an owner's review. Put spend caps, scoped keys and short credential lifetimes on every LLM API key you issue.
What Google Actually Found
GTIG found a working attack pipeline sitting on an exposed staging server. The framework, called exploit_agent, integrates several frontier and local LLMs, runs continuous differential analysis of popular web application software before and after patch announcements, drafts exploits and executes them without human intervention. Its post-compromise scripts handle internal reconnaissance, brute-forcing and webshell installation, and they also remove other attackers' webshells. Victim telemetry points at French media, entertainment and education websites.
Differential analysis, in this context, means comparing the code of a release before and after a security patch to work out which flaw the patch fixed, then writing an exploit for anyone who has not applied it yet. Defenders have long relied on the gap between a patch shipping and an attacker understanding it. A pipeline that reads every patch the day it ships removes most of that gap.
The report puts this in a wider trend. Overall disclosures and exploits nearly doubled between January and August 2026, while zero-days rose more slowly, from a 2025 average of 8 a month to 11 a month in 2026, before spiking to 22 in August. GTIG also tracks off-the-shelf multi-agent frameworks such as CAIRN, CyberStrikeAI, HexStrike AI and PentAGI, and one actor ran a customised PentAGI build against organizations in Asia from late August to early October.
Google's consumer-facing summary frames the response as "defending against machine-speed threats at human speed isn't feasible", and then sells its own AI Threat Defense product. Read that part as a vendor pitch. The threat data underneath it is the useful part.
How Fast Does an N-Day Become an Attack Now?
PaperCut is the clearest measured case, and the answer was four days. PaperCut published its advisory for CVE-2026-81578 and CVE-2026-82078 on August 27, and GreyNoise tracked an AI-orchestrated campaign from August 31 to September 2 that compromised 440 PaperCut instances at 395 organizations in 48 countries. The operator used hundreds of agents on OpenAI's Codex harness running a DeepSeek model. At least 11 organizations fell in 26 seconds once the campaign launched, and the fastest path from initial access to Domain Admin took 5 minutes.
BleepingComputer's account of the GreyNoise data adds the build time: just under four hours from an empty workspace to the first remote code execution against a real victim. GTIG separately notes that researchers and Huntress produced working exploits within a day of disclosure using LLMs.
We covered the containment side of that incident in 440 Print Servers Fell. Only 12 Reached Domain Admin.: how far the attacker got depended on what the PaperCut host's service account could reach in Active Directory. What Q4 adds is that the four-day gap is no longer one fast actor. GTIG now has a pipeline that runs the diff-to-exploit step automatically, against many products at once.
What Changed for AI Coding Agents?
A suspected DPRK cluster used a coding agent's own configuration file as the delivery mechanism. GTIG says MIDNIGHT NEPTUNE, which targets cryptocurrency firms in North America and Europe, modified .claude/settings.json in internal code repositories so that when users started Claude Code, it automatically downloaded the SOMBERMEME backdoor. The group also hid lateral movement by naming the GitHub Personal Access Tokens it generated after Microsoft Copilot.
This works because the shared project settings file is designed to travel. Anthropic's documentation says to commit .claude/settings.json so everyone who clones the repository gets the same permissions, hooks, and plugin settings, and lists hooks among the things that file is meant to carry. A hook is a shell command the agent runs at a defined point in a session. One edited JSON file in a busy repository becomes a command that runs on every developer laptop that opens it.
This is the second time we have seen the pattern. In August the ChainDrop worm used agent config for persistence, which we wrote up in npm Pulled the Packages. Your Agent Config Reinfects You.. Earlier we measured how often "docs" commits to Claude Code plugins actually changed runtime instructions in Only 16% of 'Docs' Commits Were Docs. Most review policies still classify these files as configuration or documentation, even though the agent executes what they contain.
Claude Code already ships the control. Its managed settings include allowManagedHooksOnly, which runs only the hooks your organization deploys, and disableAllHooks, which any settings file can set. The same tier can set allowManagedPermissionRulesOnly, which ignores allow, ask and deny rules from user and project files. Managed settings sit at the top of the precedence stack, and a repository file cannot loosen them. Other coding agents also read configuration from the repository, so write the review rule below to cover the folders used by Cursor and GitHub Copilot as well.
Why Are LLM Keys Now a Billing Incident?
Stolen LLM credentials are now a revenue business, and the bill lands on you. Mandiant's November 2 write-up on LLMJacking reports single-day fraudulent charges above $100,000 and cumulative charges up to $1 million before access was revoked. LLMJacking is the theft of an organization's cloud or model API access to run, or resell, inference on the victim's account. The dominant entry point was exposed or stolen cloud identity tokens: IAM credentials, service account keys, GitHub PATs and live session tokens. In several cases the LLM service had never been enabled before the compromise. The attacker enabled Bedrock or the Gemini API with the stolen key and ran up the charges within minutes or hours.
The Q4 tracker adds the resale layer. "LLM Mixers" sell pooled stolen access at $0.10 to $25 per million tokens, and attackers use a "burn and rotate" pattern, creating and deleting short-lived service accounts to stay under rate limits. In one case a botnet generated millions of API calls in about 24 hours before the credentials were revoked.
Gateways are in the path too. GTIG lists LiteLLM CVE-2026-40217 among the privilege escalation routes attackers used to turn limited Virtual Keys into administrative credentials. Obsidian Security, which disclosed the chain that CVE belongs to, says upgrading to LiteLLM v1.83.14-stable or later closes all three linked flaws. If you run LiteLLM as your gateway, it belongs in the same patch tier as your VPN. Our earlier pieces on stolen Bedrock keys from nine AI platforms and the Fireworks breach cover the rotation steps for those incidents.
The Case Against Panic
GTIG itself says these tools are noisy. Autonomous frameworks "currently trade stealth for velocity", generating forensic traffic, recognizable identifiers and distinct code-level traces, and Google SecOps already ships detection rulepacks for one of them. The report also expects stealth-focused espionage groups to stay away from them in targeted networks. Its stronger expectation is that cybercrime actors keep adopting them, because wider reach makes up for a lower success rate.
The same lower barrier has produced a flood of junk. After the October PHP flaws, CVE-2026-109618 and CVE-2026-109619, hundreds of claimed proofs of concept appeared that did not work, which GTIG says created extra verification work for defenders. Do not let every GitHub POC trigger an emergency change. Do treat the vendor advisory date as the start of the clock, whether or not a working exploit is public yet.
What to Do About It
The work is in three places: the patch clock, the repository, and the key store.
This Week:
- Pull the list of internet-facing applications from your attack surface inventory and check each one's last patch date against its vendor's latest security advisory. Anything more than a week behind goes to the owning team today.
- Search every repository for
.claude/settings.json,.claude/settings.local.jsonand hook definitions, and diff them against their last reviewed version. Any hook that downloads or runs a remote script is an incident until proven otherwise. - Run a search across your cloud audit logs for LLM services enabled in the last 90 days, using the CloudTrail, Azure and GCP event names Mandiant lists. Any enablement that did not come from your infrastructure-as-code pipeline needs an owner.
This Month:
- Rewrite the patch policy so internet-facing n-days with a public vendor fix have a remediation target in days, measured from the advisory date. PaperCut gave its customers four.
- Add agent config paths to CODEOWNERS and turn on branch protection's "Require review from Code Owners" so a security reviewer must approve any change to
.claude/or the equivalent folders for other agents. - Deploy Claude Code managed settings with
allowManagedHooksOnlyset, so only hooks your platform team ships can run on developer machines. - Upgrade any self-hosted LiteLLM to v1.83.14-stable or later and confirm the version in production, not in the repo.
Before Your Next Budget Cycle:
- Move every LLM API key into a secrets manager with automated rotation, and scope each key to a model, an environment and a spend limit, as Mandiant recommends. A leaked utility key should not be able to call your most expensive frontier model.
- Set budget alerts and per-model quotas on every cloud account that can enable an LLM service, including accounts that do not use one today.
The Bottom Line
Exploit development has followed the same path as every automated attack before it. Worms automated propagation, exploit kits automated delivery, and now LLM pipelines automate the step that used to need a skilled human reading a patch. Each time, the defenders who held up best had already shortened their own cycle when the automation arrived. GTIG's Q4 tracker is evidence that the diff-to-exploit step has been automated in the wild. Your patch clock, your agent config review and your key controls were set for a slower attacker.
Start with the repository search for .claude/settings.json hooks; it takes an afternoon.
Continue Reading
- 440 Print Servers Fell. Only 12 Reached Domain Admin.
- npm Pulled the Packages. Your Agent Config Reinfects You.
- Rotate Your Bedrock Keys If You Used These Nine AI Platforms
- Rotate Every Secret You Stored on Fireworks After Breach
- LiteLLM's October 21 Compromise Puts Provider Keys at Risk
- Claude Code Stops Asking Aug 14. Prompts Aren't Policy.