Claude Code

Google GTIG: First attack framework turns patches into exploits with no human involved

3 min read AI-generated

A North Korean group also tampered with .claude/settings.json in internal repos. Every Claude Code launch then pulled down a backdoor.

Featured image for "Google GTIG: First attack framework turns patches into exploits with no human involved"

Google Threat Intelligence Group (GTIG) published its quarterly AI Threat Tracker on November 16, and the title gives away the headline: “Agentic Threat Activity Crosses a Significant Threshold.” For the first time, GTIG has seen attackers use LLM frameworks to fully automate finding and exploiting vulnerabilities with no human in the loop. The previous edition, two months ago, said it hadn’t observed such pipelines in the wild.

exploit_agent diffs code before and after the patch

GTIG found the tool on an exposed staging server. The framework, called exploit_agent, is wired into several frontier models and locally hosted LLMs. It continuously compares the code of popular web applications before and after a patch announcement, works out the likely flaw, drafts an exploit and runs it automatically. The same server held stolen victim data plus scripts for brute-forcing, installing webshells and removing webshells left by rival attackers. Victim telemetry points to French sites in media, entertainment and education. One comment in the configuration is written in Russian.

GTIG thinks attackers like this may get more out of freshly patched n-day bugs than out of hunting zero-days, because the patch tells them where to look. The report shows how fast that has become with two chained flaws in PaperCut print servers: researchers used language models to build working exploits within a day of disclosure.

Claude Code as the way in, inside your own repo

If you use Claude Code, another section matters more. A suspected North Korean cluster GTIG tracks as MIDNIGHT NEPTUNE went after crypto companies in North America and Europe. It got in through fake recruiter profiles whose photos carried Google’s SynthID watermark, so they were AI-generated. Targets were sent to fake meeting sites with “ClickFix” lures and ran a supposed diagnostic script themselves.

From there the attackers moved into internal code repositories and edited .claude/settings.json so that Claude Code downloaded the SOMBERMEME backdoor every time someone started it. GTIG calls it a supply chain attack inside the organization. To stay hidden, the group also created GitHub personal access tokens named after Microsoft Copilot.

Review your project settings like code

The exploit agent makes the headline. For day-to-day work, the second case matters more. A .claude/settings.json in a repo can define commands that run when an agent starts, so anyone with write access to the repo can reach every machine where someone opens Claude Code in it. Changes to that file belong in the same review as any other code change, and an unexpected hook in it is a red flag. On the same day, Anthropic cut its vulnerability disclosure deadline to 30 days for a related reason: models now find bugs so quickly that none stay secret for months.

Sources

GoogleSecurityClaude CodeThreat IntelligenceAgents