{
  "type": "NewsArticle",
  "url": "https://sonomos.ai/newsroom/google-gtig-first-fully-automated-exploit-agent-no-human-in-loop/",
  "slug": "google-gtig-first-fully-automated-exploit-agent-no-human-in-loop",
  "title": "Google says it found an exploit framework that finds and attacks flaws with no human in the loop",
  "summary": "Google's Threat Intelligence Group says it found an exposed server running exploit_agent, which compares patched and unpatched code, drafts exploits and runs them automatically, the first fully automated pipeline it has observed.",
  "section": "Cybersecurity",
  "category": "cybersecurity",
  "format": "analysis",
  "tags": [
    "AI agents",
    "Prompt Injection",
    "LLMjacking",
    "Claude Code",
    "Supply chain attack"
  ],
  "published_at": "2026-11-21T14:12:00.000Z",
  "updated_at": "2026-11-21T14:12:00.000Z",
  "word_count": 799,
  "publisher": {
    "name": "Sonomos Newsroom",
    "url": "https://sonomos.ai/newsroom/",
    "standards": "https://sonomos.ai/newsroom/standards/",
    "organization": {
      "name": "Sonomos",
      "url": "https://sonomos.ai/"
    }
  },
  "follows_up": null,
  "followed_by": [],
  "sources": [
    {
      "number": 1,
      "title": "GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold",
      "publisher": "Google Cloud Threat Intelligence",
      "url": "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold",
      "published": "2026-11-16",
      "accessed": "2026-11-21"
    },
    {
      "number": 2,
      "title": "AI Update: Google Finds AI Tool Exploiting Software Flaws With No Human in the Loop",
      "publisher": "ID Tech",
      "url": "https://idtechwire.com/ai-update-google-finds-ai-tool-exploiting-software-flaws-with-no-human-in-the-loop/",
      "published": "2026-11-20",
      "accessed": "2026-11-21"
    }
  ],
  "verification": {
    "method": "automated-source-check",
    "checked_at": "2026-11-21T14:12:00.000Z",
    "description": "https://sonomos.ai/newsroom/standards/",
    "claims": [
      {
        "claim": "GTIG found an exposed staging server hosting exploit_agent, which compares code before and after patches, drafts exploits and executes them automatically, and describes it as the first fully automated discovery-and-exploitation pipeline it has observed.",
        "source": 1,
        "source_url": "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold"
      },
      {
        "claim": "Mandiant responded to LLMjacking incidents in 2026 where single-day fraudulent charges exceeded $100,000 and cumulative abuse reached $1 million in the most severe cases.",
        "source": 1,
        "source_url": "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold"
      },
      {
        "claim": "GTIG says a suspected DPRK cluster modified .claude/settings.json in internal repositories so that Claude Code downloaded the SOMBERMEME backdoor when a user started it.",
        "source": 1,
        "source_url": "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold"
      },
      {
        "claim": "Zero-day exploits averaged 8 a month in 2025 and 11 a month in 2026, with 22 in August 2026.",
        "source": 1,
        "source_url": "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold"
      }
    ]
  },
  "corrections": [],
  "body_markdown": "Google's Threat Intelligence Group (GTIG) said in a [report published on November 16](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold) that it found an exposed staging server running a framework called `exploit_agent`, which compares popular web applications' code before and after a patch, identifies the flaw, drafts an exploit and runs it without a person involved. GTIG describes it as the first time it has observed threat actors fully automating vulnerability discovery and exploitation.\n\nThe same report sets out how attackers are using AI agents in other parts of an intrusion: stolen access to AI models resold for profit, and a coding assistant's settings file altered to deliver a backdoor. [ID Tech's summary of the report](https://idtechwire.com/ai-update-google-finds-ai-tool-exploiting-software-flaws-with-no-human-in-the-loop/) covers the same findings.\n\n## What did Google find on the exposed server?\n\nGTIG says the server hosted a multi-agent framework that works with several frontier and local language models. It runs continuous differential analysis of popular web applications, drafts exploits from the differences and executes them automatically. After a compromise its scripts handle internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells.\n\nTelemetry suggests the operator targeted French websites in media, entertainment and education, and data was exfiltrated from victims. GTIG's worked example is a flawed access check in version 2.19.25 of the Spectra (Ultimate Addons for Gutenberg) WordPress plugin. A configuration file contained Russian-language comments, one of which tells the user to enter their LLM API keys for \"auto-exploitation\".\n\n## Why does patch diffing matter?\n\nA patch announces where a flaw was. GTIG says actors may prefer already-patched (\"n-day\") vulnerabilities because the research scope is finite: source differences, release notes and proof-of-concept code. It reports that disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose more modestly, from an average of 8 a month in 2025 to 11 a month in 2026, with 22 in August.\n\n| Case in the report | What GTIG says |\n|---|---|\n| PaperCut, late August 2026 | Two flaws (CVE-2026-81578, CVE-2026-82078) chained to give unauthenticated remote code execution. GreyNoise reported a likely LLM-assisted campaign four days after disclosure that hit hundreds of organisations and reached domain administrator within minutes. |\n| PHP, October 2026 | Flaws (CVE-2026-109618, CVE-2026-109619) found with LLM-enabled analysis. A proof-of-concept video and early patched code appeared before the October 21 patches; afterwards hundreds of claimed proofs of concept, likely LLM-generated, did not work. |\n| `exploit_agent` | Fully automated differential analysis, exploit drafting and execution against web applications. |\n\nGTIG also notes a cost to defenders: the flood of non-working, LLM-generated exploits adds verification work.\n\n## How are attackers using AI tools and access?\n\nGTIG's report covers several data and access issues that bear on organisations using AI services:\n\n- **LLMjacking.** Mandiant responded to several 2026 incidents in which financially motivated actors compromised cloud environments to consume compute or model quotas. GTIG says billing spikes can occur within hours, single-day fraudulent charges exceeded $100,000, and in the most severe cases cumulative abuse reached $1 million before remediation. In one case an actor exploited CVE-2026-40217, a sandbox escape in LiteLLM that elevates limited virtual keys to full administrative credentials.\n- **Resale of stolen access.** The report describes \"LLM mixers\", storefronts that spread queries across pools of stolen credentials, advertised on Russian- and Chinese-language underground markets at $0.10 to $25 per million tokens.\n- **Coding-assistant configuration.** In intrusions at cryptocurrency organisations in North America and Europe, GTIG says a suspected North Korean cluster it calls MIDNIGHT NEPTUNE pivoted into internal code repositories and modified `.claude/settings.json` so that [Claude Code](https://sonomos.ai/blog/is-claude-code-safe-sensitive-code-2026/) downloaded the SOMBERMEME backdoor whenever a user started it. The actor also created GitHub personal access tokens named after Microsoft Copilot to mask lateral movement.\n\n## What does Google say it is doing?\n\nGTIG says it disabled associated assets, updated products to prevent misuse and used the findings to strengthen Gemini's classifiers. Google SecOps deployed rule packs to detect ARTEX, one of the open-source offensive frameworks the report names, and Mandiant released a hardening guide. GTIG assesses it is highly likely such frameworks will be used globally, and says they trade stealth for speed because unrestricted autonomous execution leaves forensic traffic and code traces.\n\n## What is unresolved?\n\nThe report does not name the operator or say how many sites `exploit_agent` compromised beyond the French targets suggested by telemetry. It does not say which language models the framework used. GTIG's own assessment is that stealth-focused espionage actors are less likely to use noisy autonomous tools inside targeted networks, so how widely the fully automated approach spreads is a forecast, not a measurement.\n\nEarlier in this newsroom, [OpenAI's reports of test models bypassing access controls](https://sonomos.ai/newsroom/openai-reports-test-models-bypassed-access-controls/) covered the opposite problem: agents exceeding their authority without any attacker. For background on how agents handle data, see [how AI agents touch private data](https://sonomos.ai/blog/ai-agents-data-privacy-2026/).",
  "links": {
    "html": "https://sonomos.ai/newsroom/google-gtig-first-fully-automated-exploit-agent-no-human-in-loop/",
    "markdown": "https://sonomos.ai/newsroom/google-gtig-first-fully-automated-exploit-agent-no-human-in-loop/index.md",
    "json": "https://sonomos.ai/newsroom/google-gtig-first-fully-automated-exploit-agent-no-human-in-loop/index.json",
    "image": "https://sonomos.ai/og/newsroom/google-gtig-first-fully-automated-exploit-agent-no-human-in-loop.png"
  }
}
