Google's Threat Report: Agents That Diff Patches Into Exploits, and a Backdoor Hidden in Claude Code's Settings

On 16 November 2026 Google Threat Intelligence Group (GTIG) and Mandiant published their quarterly AI Threat Tracker, titled “Agentic Threat Activity Crosses a Significant Threshold”. Most of it extends trends we have covered before, such as attackers adopting open-source agent pentest frameworks like ARTEX. Two findings are new, and both matter to anyone who writes or ships software:

  1. exploit_agent: a multi-agent framework that compares code before and after a patch, finds the bug the patch fixed, writes an exploit and runs it, all without a human. GTIG calls this the first time it has seen discovery and exploitation fully automated by attackers.
  2. A coding agent’s configuration used as a supply-chain vector. A suspected North Korean cluster edited .claude/settings.json inside victims’ internal repositories so that starting Claude Code downloaded a backdoor GTIG calls SOMBERMEME.

SecurityWeek’s write-up of the same report independently confirms the main points and adds one detail: the North Korean cluster, MIDNIGHT NEPTUNE in Google’s new naming scheme, is the group Google previously tracked as UNC1069.

exploit_agent: patch diffing without a human

GTIG found exploit_agent on an exposed staging server, which is the only reason anyone outside the operator knows it exists. According to the report, the framework connects to “multiple frontier and local LLMs” and is set up for continuous differential analysis of popular web application software. Whenever a patch is announced, it compares the vulnerable and fixed code, works out what changed, drafts an exploit and runs it against targets automatically.

The report publishes one artefact. It is a PHP function from the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25, that the framework was examining. The function blocks access to a REST route only when the request URI contains /wp/v2/spectra-popup. Checking an authorization rule with a substring match on the raw URI is the kind of logic flaw a model can reason about cheaply once a diff points it at the right file. GTIG also prints comments from the framework’s environment file. They are in Russian and translate as “LLM API keys for auto-exploitation (exploit_agent pilot)” and “Enter the keys yourself — right here, not in the chat.” The second line suggests the operator built the tool with a chat assistant. GTIG does not attribute exploit_agent to any group, and Russian comments alone do not settle who ran it.

The same server also held data stolen from victims and scripts for internal reconnaissance, brute-forcing, installing webshells, and removing other attackers’ webshells. Victim telemetry pointed to French media and entertainment sites and to the education sector.

Why this matters. For a long time the main defence against n-day attacks was that turning a patch into a working exploit took a skilled person hours or days. GTIG’s argument is that n-days suit LLMs well, because the search space is small: one diff, the release notes, perhaps a proof of concept. Zero-day hunting is wider open-ended work. The report’s numbers fit this. Disclosures and exploits “nearly doubled” between January and August 2026. Zero-days exploited in the wild rose much less, from an average of 8 a month in 2025 to 11 a month in 2026, before a spike to 22 in August.

The report gives two concrete cases of how short the window has become:

  • PaperCut (late August 2026): two bugs, CVE-2026-81578 (authentication bypass) and CVE-2026-82078 (unsafe class loading), chained into unauthenticated remote code execution. Researchers and Huntress produced working exploits with LLM help within a day of disclosure. GreyNoise saw a campaign four days after disclosure that likely used LLMs to build the exploit and the target list, and went from foothold to domain admin “within minutes”.
  • PHP (CVE-2026-109618/109619, patched 21 October): a proof-of-concept video and early patched code leaked before the official fix. After the patch, hundreds of claimed proofs of concept appeared that did not work, probably generated by LLMs, and defenders had to spend time checking each one. Cheap exploit generation produces noise as well as real attacks.

The Claude Code supply-chain pivot

The North Korean case is short in the report, but it is the part developers should read most carefully. According to GTIG, MIDNIGHT NEPTUNE went after cryptocurrency organisations in North America and Europe as follows:

  1. Fake recruiters. Profiles on professional networks used photos carrying SynthID watermarks, Google’s watermark for AI-generated media, which GTIG takes as proof the personas were synthesised.
  2. ClickFix. Targets were sent to fake meeting sites that told them to run a “diagnostic fix” script, which ran a staging script that was probably also written by an LLM.
  3. The repository pivot. Once inside, the operators modified .claude/settings.json in internal code repositories, so that when colleagues started Claude Code it automatically downloaded SOMBERMEME.
  4. Camouflage. They created GitHub personal access tokens named after Microsoft Copilot, so the tokens looked like normal tooling.

No new vulnerability in Claude Code was involved. .claude/settings.json is the shared, committed project-settings file, and the hooks it defines are meant to run commands on events such as SessionStart. Teams use them for formatters, test runners and environment setup. Claude Code’s documentation gates some repository-supplied settings behind a workspace-trust dialog, but that trust is stored per repository, keyed on the git root, not per revision of the file. A developer who trusted their company’s monorepo months ago has, in effect, already agreed to run whatever hooks it defines. The documentation also states that hooks from settings files are used in claude -p and SDK runs even in folders that were never trusted, which matters for CI jobs that run agents over repositories. One malicious merge to .claude/ runs code on every engineer’s machine that pulls it.

GTIG did not publish hashes, domains or command-and-control details for SOMBERMEME, and it names no victims. Anthropic is not quoted in the report. The idea itself is not new: security researchers have warned for most of this year that repository-supplied agent configuration is an attack surface. What is new is a state-backed group using it after a targeted intrusion, as a quiet way to move laterally inside a company.

The rest of the tracker, briefly

  • Agent frameworks in live intrusions: beyond ARTEX, GTIG saw CAIRN (an OODA loop where agents share a “blackboard”; used against US aerospace targets), PentAGI (used late August to early October 2026 for scanning across Asia), and experimentation with CyberStrikeAI and HexStrike AI. HexStrike is an MCP-based harness exposing more than 150 offensive tools to an LLM client. GTIG’s view is that these tools currently “trade stealth for velocity”: they are noisy and leave traces, but they still achieve their objectives.
  • LLMJacking: Mandiant saw fraudulent cloud AI charges above $100,000 in a single day, and around $1 million in total in the worst cases. One actor used CVE-2026-40217, a LiteLLM sandbox escape that turns limited “virtual keys” into admin credentials. Stolen access is resold through “LLM mixers” at $0.10 to $25 per million tokens.
  • Influence operations: one contractor was building a platform to run more than 100,000 accounts on cloud phone farms, with LLMs simulating normal scrolling behaviour so new accounts look aged.

What to take from it, and what to doubt

This is a threat-intelligence report from a vendor that also sells the defences (Google AI Threat Defense, Gemini 4 Carbon’s cyber safeguards). Its counts come from Google’s own telemetry and cannot be checked from outside. “First time without a human in the loop” means the first time GTIG has observed it, and it found this one only because a staging server was left open. The report also undercuts some of the hype itself: many AI-generated exploits do not work, and the noise these tools make is currently a weakness for attackers.

The practical lessons for engineering teams do not depend on those caveats:

  • Treat agent configuration as executable code. Put .claude/, .cursor/, .mcp.json and similar paths under CODEOWNERS with the same review as .github/workflows/, and alert on changes to hooks, env and apiKeyHelper.
  • Review what your CI agents load. Headless agent runs over untrusted branches are where repository hooks run with no prompt at all.
  • Assume n-day exploits arrive within days. For internet-facing software, a monthly patch cycle is now too slow, and a public diff is effectively a public exploit.
  • Watch for oddly named tokens and AI spending spikes. A PAT named “Copilot” or a sudden jump in model usage can be an early sign of compromise.

Sources