On November 16 Google Threat Intelligence Group (GTIG) published its latest AI Threat Tracker on its official blog and X. It says that last quarter it saw threat actors use LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time: on an exposed staging server it found exploit_agent, a multi-agent framework connected to several frontier and local models that continuously diffs popular web software before and after patches, infers the flaw, drafts an exploit and runs it automatically. The report also says a suspected MIDNIGHT NEPTUNE cluster modified .claude/settings.json files in internal repositories so that starting Claude Code downloaded the SOMBERMEME backdoor.
Key Takeaways
- ✓exploit_agent: a multi-agent vulnerability research and weaponization framework connected to several frontier and local LLMs that diffs code before and after patches, identifies the flaw, drafts an exploit and runs it with no human involved
- ✓The same server held exfiltrated victim data plus scripts for internal reconnaissance, brute-forcing, installing webshells and removing rival attackers' webshells; victim telemetry pointed to French media, entertainment and education websites
- ✓The configuration file had Russian-language comments; GTIG did not attribute a group and gave no CVE, victim count or success rate
- ✓A suspected MIDNIGHT NEPTUNE cluster modified .claude/settings.json so starting Claude Code downloaded the SOMBERMEME backdoor, and created GitHub personal access tokens named after Microsoft Copilot to mask lateral movement
- ✓The report also names multi-agent offensive frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI, and LLMJacking of cloud LLM quotas
Key Decision Metrics at a Glance
Heavy Claude Code use: compare subscription limits and API bills
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
Patch diffing is an old attacker technique, but it used to take a skilled person days. Google Threat Intelligence Group's quarterly AI Threat Tracker, published on its official blog and X on November 16, says agents have now taken over that step, shrinking the window between a patch and a weaponized exploit.
How it works
GTIG found exploit_agent on an exposed staging server: an orchestrator coordinating specialized agents, connected to several frontier and local models, that continuously compares popular web software before and after patches, infers the flaw, drafts an exploit and runs it automatically. The server also held victim data and scripts for reconnaissance, brute-forcing and webshells. In a separate case, attackers modified .claude/settings.json so that starting Claude Code downloaded the SOMBERMEME backdoor.
Numbers that matter
There are no benchmarks, and the report gives no CVE, victim count, success rate or patch-to-exploit time. GTIG notes that many LLM-generated exploits claim to work but do not, adding verification work for defenders.
What it means for developers
GTIG's main recommendation is to shorten patch windows. Teams using AI coding tools should put agent configuration files such as .claude/settings.json under code review and look for hooks or commands nobody added, and should regularly audit GitHub personal access tokens and cloud LLM quota usage.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.