Google: AI agent turned patches into exploits and stole data
Original title: Google: AI agent turned a security patch into a working exploit with no human help
GTIG says a Russian-language framework autonomously weaponized patches, then stole data from French media, entertainment, and education sites.
Google Threat Intelligence Group says a suspected Russian-speaking operator used a multi-agent framework, exploit_agent, to diff security patches, build exploits, and run them with no human in the loop. An exposed staging server also held stolen victim data and post-compromise scripts for reconnaissance, brute-forcing, and webshells. GTIG says other actors, including suspected PRC-nexus groups, are testing AI offensive frameworks such as PentAGI, ARTEX, CAIRN, CyberStrikeAI, and HexStrike AI.
Unconfirmed: The operator’s Russian-language affiliation and the claim this is the first fully autonomous case are GTIG assessments.
- The framework compared code before and after updates in popular web software.
- Victim telemetry pointed to French media, entertainment, and education sites.
- PentAGI was also seen with proxy services against the aerospace sector.
- GTIG calls fully autonomous discovery and exploitation a first in its Q4 2026 tracker.
Impact: Stolen data plus webshells and internal access
| Product | Affected |
|---|---|
| popular web application software | — |
- 1.Hunt for webshells and unexpected post-compromise scripts on exposed web apps.
- 2.Review whether AI offensive frameworks are being used against public-facing apps.
- Vendors
- Malware
- exploit_agent
- Organizations
- Google Threat Intelligence Group
From the original article
2 min read at theregister.comA suspected Russian-speaking operator used an AI agent to turn security patches into working exploits with no human in the loop – a first such occurrence, according to the Google Threat Intelligence Group (GTIG) – before ultimately stealing data from victim organizations.
By Jessica Lyons · theregister.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.