Up to date as of November 21, 2026. This article draws on Google’s report of November 16, 2026 and on the FBI alert of November 13, 2026. The tools described are evolving fast: it will be expanded as new publications come out.
Between the release of a patch and its installation, a vulnerability is known to everyone but still open. Analyzing a patch to derive a working attack from it used to take time and skill. AI agents are automating that step, and shrinking the time defenders have accordingly.
What Google observed
On November 16, 2026, Google (Google Threat Intelligence Group and Mandiant) published a new review of offensive uses of AI. For the first time, its teams observed attackers fully automating the discovery and exploitation of vulnerabilities, with no human involved:
- on a server left exposed, a set of agents dubbed exploit_agent, connected to several language models, compared the code of popular web applications before and after patch announcements, identified the fixed flaw, drafted an exploit and launched it automatically. The example analyzed by Google concerned the access controls of a WordPress plugin;
- the same server held scripts to brute-force passwords, install web backdoors (“webshells”) and remove those of other attackers;
- according to victim data, the operator was specifically targeting French websites in media, entertainment and education.
Google also notes that vulnerability disclosures and exploits nearly doubled between January and August 2026. Flaws exploited before any patch (“zero-days”) rose from an average of 8 per month in 2025 to 11 per month in 2026, peaking at 22 in August. Another effect: after every major patch, hundreds of supposed AI-generated proofs of concept, often non-functional, force defenders to sort the real from the fake.
Autonomous attack tools
On November 13, 2026, the FBI published an alert about ARTEX, an open-source penetration testing platform driven by language models and hijacked by attackers:
- since at least August 2026, it has been used against organizations in many sectors: finance, government, technology, defense, energy, healthcare, manufacturing, telecommunications, legal;
- some intrusions went from initial targeting to privileged access in under three hours;
- the original repository was taken down in October 2026, but the tool keeps circulating through copies and a commercial version. In early November, more than 1,600 ARTEX consoles were exposed on the internet.
Google observes that these tools currently trade stealth for speed: they leave network and code traces that make them easier to detect. It expects them to become stealthier.
Finally, attackers are going after access to AI itself. In 2026, Mandiant handled several cases of “LLMjacking”: cloud environments compromised to consume the victim’s AI model quotas. Fraudulent charges exceeded 100,000 dollars in a single day in some cases, and reached one million dollars in total in the most severe ones.
What changes for defenders
The methods are not new: compromised credentials, unpatched systems. What changes is the speed and the scale. An internet-facing application that is not patched quickly after a fix is released becomes a target, including for inexperienced attackers.
In France, the plan to strengthen cybersecurity presented by the government on November 19, 2026 plans to mobilize 200 million euros from France 2030 to make AI an asset for cyber defense: faster detection of attacks, identification of vulnerabilities and automation of some defensive actions. The race is being run on both sides.
Shortening the time to patch
Know what is exposed
You cannot patch what you do not know about. The FBI recommends inventorying every internet-facing application, including the most peripheral ones: partner and supplier portals, the backends used by mobile apps, legacy business applications, third-party integrations, sites created by individual departments. CMS plugins, such as WordPress ones, are part of the inventory.
Prioritize according to actual exploitation
The theoretical severity of a flaw is not enough to set priorities. The urgent patches are those for flaws that are exposed on the internet and actively exploited, flagged for example by the US CISA catalog of exploited vulnerabilities or by CERT-FR alerts. Written deadlines, shorter for these cases, prevent everything from being handled at the same pace.
Protect while waiting for the patch
When a patch cannot be applied right away, reduce the exposure:
- restrict administration interfaces (cloud, network devices, backup, application publishing) to trusted networks or dedicated access paths, as the FBI recommends;
- put internal applications behind “zero trust” access rather than publishing them directly;
- disable unnecessary features and plugins, and apply a temporary filtering rule on the web application firewall;
- document the accepted risk and the planned patch date.
Detect while these tools are noisy
Automated attacks currently leave traces: bursts of requests, chains of tests, creation of web backdoors, new accounts or new secrets added to applications. These signals must reach the SIEM, with dedicated detection rules (see our analysis Logs, the foundation of the SOC). The FBI also recommends monitoring credential additions, privilege assignments and unusual activity by application accounts.
Protect access to AI services
Against LLMjacking: continuously search code, pipelines and container images for forgotten secrets, and rotate them without delay; limit the rights of service accounts; set quotas and consumption alerts on AI services, because a sudden spike in usage is often the first sign.
What the texts say
- NIS2: Article 21 includes vulnerability handling among the mandatory measures. For the digital providers it covers, Implementing Regulation 2024/2690 specifies (point 6.6 of the Annex) that security patches must be applied “within a reasonable time”, tested before production and verified, and that additional measures are required when a patch is not applied. A good reference for all entities (see NIS2 in France).
- Cyber Resilience Act: since September 11, 2026, manufacturers of products with digital elements must report actively exploited vulnerabilities (see One incident, several clocks).
Where to start
- Inventory everything exposed on the internet, including peripheral applications and plugins.
- Measure your actual delay between the release of a patch and its deployment, by type of system.
- Set written deadlines, shorter for exposed and actively exploited flaws.
- Follow the alert sources: CISA catalog, CERT-FR alerts, vendor advisories.
- Plan the interim measures: restricted access, application filtering, disabling unnecessary features.
- Detect automated attacks in the SIEM and prepare the response.
- Protect secrets and access to AI services, with consumption alerts.
AI also helps defenders, provided it is properly governed (see The AI-augmented SOC). And on phones, attacks already exploit flaws that were fixed but never installed (see The smartphone, security’s blind spot).
Sources
- Google Threat Intelligence Group and Mandiant, GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, November 16, 2026
- FBI, FLASH-20261113-001, Threat Overview of ARTEX Autonomous Penetration Testing Framework, November 13, 2026
- French government, plan to strengthen cybersecurity, November 19, 2026 (in French)
- CISA, Known Exploited Vulnerabilities (KEV) catalog
- CERT-FR, security alerts (in French)
- Implementing Regulation (EU) 2024/2690 (NIS2 technical requirements), EUR-Lex
- Directive (EU) 2022/2555 (NIS2), EUR-Lex
- Regulation (EU) 2024/2847 (Cyber Resilience Act), EUR-Lex
