Slashgear Google Reports AI Agents Automating Cyberattacks, Including Exploit Development From Patch Diffs
Article Content
- •GTIG reports threat actors are using multi-agent AI frameworks (including Exploit Agent) to automate vulnerability discovery and exploit development.
- •Exploit Agent reportedly compared pre- and post-patch code to find flaws and deployed exploits without approval.
- •GTIG observed an AI-driven credential theft campaign that was planned and executed within six hours in Q2 2026.
- •LLMJacking, stealing AI credentials and cloud keys for resale, is reported on underground marketplaces.
- •Defenders are advised to rely on network-level blocking such as virtual patching, not only on patch deployment.
Google's Threat Intelligence Group (GTIG) warned that threat actors are using AI agents and multi-agent frameworks, including ARTEX, CAIRN, CyberStrikeAI, and PentAGI, to automate stages of cyberattacks such as finding and exploiting vulnerabilities. GTIG described a multi-agent framework called "Exploit Agent" (reported as consisting of Russian-language tools) that compared code before and after a security patch, developed exploits, and deployed them without approval, with one host reportedly targeting French entertainment, media, and educational websites. Slashgear reports that over 20 zero-day vulnerabilities were exploited in August 2026, compared with an average of eight per month a year earlier. GTIG also reported a Q2 2026 case in which an attacker used AI agents to plan, build, and execute a large-scale credential theft campaign within six hours. State-linked activity is reported: Iranian group CALANQUE ION reportedly used Gemini in social engineering and reconnaissance, and China-backed hackers reportedly used an AI model to store malicious files and exfiltrate documents. Attackers also practice "LLMJacking," stealing AI credentials or cloud API keys via social engineering or exposed GitHub repositories and reselling access, with listings reportedly priced from $0.10 to $25 per million tokens on Russian- and Chinese-language marketplaces. GTIG also identified zero-day exploit code it presumes was AI-written, based on features such as nonexistent severity scores. Experts cited by Jkn News argue that patch-centred defence has limits and that network-level blocking, such as virtual patching from F5, is needed. Neither article names specific CVEs for the exploit campaigns described.
Ask AI about this cluster
Answers cite the sources they use
Timeline
More articles in this cluster (2)
Following this threat?
Track Calanque ION and F5 in your own feed — you're alerted when they show up in new reporting, leak sites or exploitation.
Free account · no card needed
Common questions
Am I affected by these AI-driven attacks?
How urgent is this?
What should defenders do first?
Continue Reading
Google Reports First Fully Autonomous AI Cyberattack Using exploit_agent Google's Threat Intelligence Group (GTIG) has identified the first instance of an autonomous AI framework, named exploit_agent, that can fully automate the process of discovering and exploiting software vulnerabilities without human intervention. This framework continuously compares code before and after security…