Google has documented the first fully autonomous AI attack tool it has seen in the wild: exploit_agent. The framework compares the code of popular web applications before and after security updates, infers vulnerabilities, writes exploits and runs them without a human in the loop. For anyone who operates websites, plugins or web applications, an old calculation is shifting: the time between a patch and an attack is shrinking.
What the Google Threat Intelligence Group found
The finding appears in the current AI Threat Tracker from the Google Threat Intelligence Group (GTIG), published on 16 November 2026. GTIG and Mandiant discovered an exposed staging server hosting a multi-agent framework for vulnerability research and weaponisation. The attackers themselves call it exploit_agent. It is wired to several frontier models as well as locally run language models.
According to GTIG, the system is set up for continuous differential analysis of popular web application software. It compares code before and after a patch is announced, analyses the changes for potential vulnerabilities, drafts matching exploits and executes them automatically. GTIG writes that this is the first time it has observed attackers fully automating vulnerability discovery and exploitation with LLM frameworks, with no human in the loop.
exploit_agent and WordPress: a look inside a plugin
One example in the report is likely to catch the attention of many WordPress operators. GTIG shows how the framework inspects access control in the Spectra plugin (Ultimate Addons for Gutenberg) at version 2.19.25. Specifically, the agent analyses a function that is meant to allow REST API access to popup content only for logged-in administrators. The plugin in the WordPress directory reports more than one million active installations and currently stands at version 2.20.6.
The distinction matters: the report shows that exploit_agent examined this code. It does not say whether an exploitable flaw was found. The example mainly illustrates the method: the tool looks for permission checks, the exact places where a patch closes a hole and thereby reveals where it used to be.
On the server, GTIG also found data exfiltrated from victims and scripts for the period after a compromise. These include internal reconnaissance, brute-force attacks, installing webshells and removing other attackers’ webshells. Victim telemetry on the server suggests, according to GTIG, that the operator was particularly interested in French websites in media, entertainment and education. Comments in a configuration file are in Russian, including a note to enter the API keys for “auto-exploitation” there and not in chat.
Why the patch window is shrinking
The logic is simple. Anyone hunting an unknown flaw (a zero-day) has to search widely. Anyone wanting to exploit already patched flaws (n-days) has a finite set of material: two code versions, release notes, advisories and, sometimes, published proofs of concept. According to GTIG, language models can now largely take over this work, including code from pre-release or preview channels.
GTIG also cites figures. The number of disclosed and exploited vulnerabilities almost doubled between January and August 2026. Zero-day exploits, by contrast, rose only modestly, from an average of 8 per month in 2025 to 11 per month in 2026, before jumping to 22 in August. As an example of the speed involved, the report points to two flaws in the PaperCut print management software (CVE-2026-81578 and CVE-2026-82078). Researchers used language models to produce working exploits within a day of disclosure, and four days later a broad campaign against hundreds of organisations began.
There is a downside. According to GTIG, the lower barrier to entry has also produced a wave of LLM-generated exploits that are claimed to work but do not. That creates extra verification work for defenders, who have to tell genuine proofs of concept from fake ones. We described the same pattern earlier in the patch gap around GLM 5.3 and Firefox, though in a research setting rather than in the wild.
Agentic attack frameworks have become routine
exploit_agent is not the only case in the report. GTIG observes attackers using open-source pentesting frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI in live intrusions. It also mentions HexStrike AI, a tool that exposes more than 150 security utilities to language models via the Model Context Protocol. What these tools have in common is that they find low-complexity vulnerabilities, generate single-use payloads and adjust their tactics based on feedback. We reported in October in more detail on the use of ARTEX by the China-nexus group VAPOR PANDA.
A side issue: LLMJacking and stolen API keys
Another section of the report is relevant for development teams. Under the term LLMJacking, GTIG describes how criminals steal credentials for cloud and AI services and consume or resell their victims’ compute quotas. Mandiant handled several such cases in 2026. In some, fraudulent charges exceeded 100,000 US dollars in a single day; in the most severe, cumulative abuse reached one million US dollars before remediation. In one case, a botnet using legitimate, probably leaked credentials generated millions of API calls in roughly 24 hours. Anyone keeping API keys in repositories, environment files or CI systems should take budgets, alerts and key rotation seriously.
What web developers and operators should do now
The report does not describe a new flaw that can be closed with a single update. It describes a speed problem. A few sober consequences follow:
- Stop batching security updates: anyone who leaves plugin or framework patches for the next maintenance window gives automated attackers days they no longer need. Automatic updates for security releases are, for many sites, more sensible than waiting to apply them by hand.
- Reduce the attack surface: every unused plugin and every open REST endpoint is a candidate for differential analysis. What is not needed should be disabled and removed.
- Watch for webshells: because exploit_agent comes with scripts for installing webshells, integrity checks of files in upload and plugin directories are worthwhile.
- Treat proofs of concept critically: not every published exploit works. For prioritisation, vendor advisories and confirmed exploitation matter, not the number of repositories.
For developers of their own plugins there is a further point: security fixes should appear as completely as possible and in a single step. As the report shows, even a partial fix published early can be enough for attackers to run the comparison. That is not an argument against transparency, but it is one for coordinated releases. A recent example of how quickly a known plugin flaw can become relevant was the upload flaw in Contact Form 7.



