Security researchers at Google warn that the rise of large language models (LLMs) is opening the door to new, industrial-scale cyberattacks. In one reported case, attackers used an LLM to discover a previously unknown vulnerability that allowed them to take control of a widely used web administration tool; the researchers believe a criminal actor intended to scale the technique, and the discovery prevented a broader assault.
What threats does Google describe?
The Google team cataloged several concrete ways LLMs are accelerating and easing cyberattacks:
-
Morphing malware: LLMs can generate malware that evades detection by altering parts of its code on each replication or infection. Such programs may rewrite their decryption routines, replace commands with alternative sequences that produce the same result, or add nonfunctional subroutines, all of which hinder antivirus detection while preserving the malicious payload for data theft, backdoor installation, or file encryption.
-
Identifying logical flaws: Unlike many traditional security tools that detect known patterns or fuzz code with random inputs, LLMs can reason about what code is intended to do and use that reasoning to find logical vulnerabilities. These flaws can remain invisible to standard automated tools and would normally require focused human review.
-
Obfuscation networks: Threat actors commonly assemble ad hoc chains of routers, servers, and specialized tools to hide origins and bypass defenses. AI-driven tools can orchestrate malicious traffic through multiple compromised intermediary servers while avoiding patterns that would trigger typical security monitors.
-
Insecure AI infrastructure as an attack surface: AI models, tooling, and supporting software themselves are becoming attractive targets. Compromising insecure AI components can give attackers a foothold to move laterally within networks, steal data, deploy ransomware, or disrupt operations.
Related cases and research
Google’s findings align with other recent work. Anthropic has warned that its Claude Mythos Preview model can find previously unknown vulnerabilities. The cybersecurity firm Calif used that model to penetrate Apple’s security; Calif reported the exploit to Apple, which is working on a patch.
The United Kingdom–backed AI Security Institute (AISI) reported that Claude Mythos Preview and OpenAI’s GPT-5.5 could reliably execute attacks that would be expected to take humans about three hours — a marked increase from a prior forecast of one hour. By comparison, when Claude Opus 4.6 debuted it could perform attacks that would take people roughly 30 minutes. In AISI’s tests the models were limited to 2.5 million output tokens; when allowed more tokens, the models were able to execute attacks that would take human attackers longer.
Why this matters
Google’s analysis highlights a widening gap between the ability of LLMs to find security flaws and commonly used defensive methods. If next-generation models can discover and exploit bugs faster than security teams can patch them, automated, large-scale attacks become a realistic threat. The report may prompt increased government scrutiny and complicate regulatory and commercial responses, since AI functions simultaneously as a defensive tool, an offensive capability, and a target.
Expert considerations
Experts who have worked with Claude Mythos Preview confirm it represents a clear advance for both offensive and defensive security tasks. There is cautious optimism that current patches will improve network security and that lessons learned will inform safer AI rollouts. Beyond immediate fixes, software developers will need to invest more in proactive defensive research to find vulnerabilities before threat actors do.
Summary
According to Google, large language models are enabling new methods for uncovering and exploiting vulnerabilities that can be scaled up to industrial levels. The report lists specific threat patterns — morphing malware, logical-flaw detection, obfuscation networks, and targeting of AI infrastructure — and relates these to recent findings involving Claude Mythos Preview and GPT-5.5. The next challenge is faster, more proactive discovery of flaws and stronger defensive measures.
