Google's GTIG finds 'exploit_agent', the first framework that turns a vendor patch into a live exploit with no human in the loop. Zero-days hit 22 a month in August. What to change in your patch and detection plan
Google's fourth-quarter AI threat report moves the attacker's clock to the moment a patch is published, documents a backdoor delivered through a coding agent's settings file, and prices stolen API access at $0.10 per million tokens.

Google's Threat Intelligence Group published its fourth-quarter AI Threat Tracker on Monday under the title "Agentic threat activity crosses a significant threshold." The threshold is specific: for the first time, GTIG says it has identified actors using language-model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop. In September the same group had described the shift to agents as incremental. Two months later it has a named artefact.
What exploit_agent does
The framework was recovered from an exposed staging server. It integrates several frontier and local models and runs continuous differential analysis of popular web application software, comparing code before and after a patch is announced to work out what the patch fixed. It then drafts an exploit, executes it, and hands off to post-compromise scripts that run internal reconnaissance, brute-force credentials, install webshells and remove other attackers' webshells. Telemetry points to French media, entertainment and education websites as victims. GTIG gives no timing figures for the tool itself, but the related PaperCut campaign that GreyNoise tracked in September is the reference point: 440 instances at 395 organisations in 48 countries, at least 11 compromised within 26 seconds of launch, and a fastest path to domain admin of five minutes.
Around the headline finding sit the trend numbers. Vulnerability disclosures and exploits nearly doubled between January and August. Zero-days averaged eight a month in 2025, 11 a month this year and 22 in August. GTIG tracked one actor running a customised build of the open-source PentAGI platform against organisations across Asia from late August to early October, and lists CAIRN, CyberStrikeAI and HexStrike AI among frameworks in criminal and suspected state hands.
The settings.json backdoor
The report's most practical detail concerns a suspected North Korean cluster Google calls MIDNIGHT NEPTUNE, targeting cryptocurrency firms in North America and Europe. It modified .claude/settings.json in victims' internal code repositories so that starting Claude Code automatically downloaded a backdoor Google names SOMBERMEME, and it hid lateral movement by naming the GitHub personal access tokens it minted after Microsoft Copilot. The lesson generalises to every coding agent: project-level configuration that can run hooks or commands on start-up is executable code, and a cloned repository is untrusted input.
LLMjacking as a market
Mandiant's companion write-up prices the theft of model access. Single-day fraudulent charges exceeded $100,000 in some cases and reached $1 million cumulatively before access was cut; "LLM mixers" resell pooled stolen keys at $0.10 to $25 per million tokens; attackers create and delete short-lived service accounts in a "burn and rotate" pattern to stay under rate limits; and a privilege-escalation flaw in LiteLLM, CVE-2026-40217, was used to turn restricted virtual keys into administrative ones. One botnet generated millions of API calls in about a day before revocation.
The counter-argument
GTIG is careful about limits. The frameworks "trade stealth for velocity": they leave recognisable traffic, identifiers and code-level traces, and Google has shipped behavioural detections and Mandiant hardening guidance. After October's PHP flaws, hundreds of claimed proofs of concept turned out not to work. Autonomy is real, but it is noisy and often wrong, and the report does not claim that every user of PentAGI or HexStrike is an attacker.
What to do
Reorder patching around publication, not severity: the diff is the exploit's blueprint, so internet-facing web applications get patched or shielded within hours of a vendor advisory, with Project Zero's new 30-day clock as the outer bound. Hunt for the post-patch signature, a webshell drop or an unexpected admin login within 72 hours of an advisory. Put agent configuration files under the same review and integrity monitoring as CI pipelines, and bind approval of hooks to content, not file names. Cap spend per API key, alert on token-volume anomalies, and rotate anything long-lived. And use the window Google describes: detection engineering against loud frameworks now buys time against the quiet ones that follow.
Sources
BERI summary of the GTIG Q4 tracker ↗TechNode Global ↗Google Cloud Threat Intelligence: agentic threat activity crosses a significant threshold ↗The story so far
- Nov 17, 2026 Qwen4-Max at $1.50 / $4.50: 2.4 trillion parameters, 71 billion active, a 1M window, and a benchmark story that depends …
- Nov 17, 2026 macOS 27.3 tightens Full Disk Access because of AI agents: silent denials, no one-click toggle, and a 'Private Folders' …
- Nov 17, 2026 Corma's $125 million Series A at $1 billion, three months after stealth: what a 16x step-up on 25 people and undisclosed…
- Nov 17, 2026 California v. OpenAI: 80-plus loss-of-control incidents, 700 agents attacking Hugging Face, and a public-nuisance theory…
Key terms in this story
- Full X disclosure (Project Zero) Google Project Zero's 2026 policy: a vulnerability it reports is published X days later whether or not a fix exists; X is 90 now, 60 from February 2027 and 30 from January 2028.
- Rate limit The maximum requests or tokens per minute a vendor allows an account, often tiered by spending history.
- Webshell A small script an attacker uploads to a web server so they can run commands on it through a browser; the classic first foothold in a server intrusion.
- Token The unit models read and write; roughly three-quarters of an English word. Pricing and context limits are counted in tokens.
- Agent An AI system that pursues a goal by taking actions with tools, in a loop, rather than answering a single question.
More in Security
All →
Java4Shell (CVE-2026-114938): an out-of-bounds write in the Java runtime's Base64 decoder, patched Monday, no exploitation seen yet. What to patch, in what order, and how to prove it
The bug is in the JVM, not a library, and it runs before your code gets a vote; that makes it an inventory problem first and a patching problem second…

Microsoft's agent-framework package was poisoned on PyPI for 45 minutes with a stealer that ran at every Python start-up. Two versions, one webhook, at most 50 victims, and a checklist for this morning
The attack did not need you to import the package. A .pth file in site-packages ran the stealer whenever any Python interpreter launched, which is why…

Project Zero's 'Full X' policy: 90 days now, 60 from February, 30 from January 2028, fixed or not. What LLM-speed exploitation means for your patch SLAs
Google's bug-hunting team has scrapped the 30-day soak period it added in 2021 and will publish vulnerability details on a fixed clock because, it say…
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.