Himalaya Harbinger, Rudrapur Burea
In the latest edition of its artificial intelligence (AI) threat report released on Tuesday, Google said hackers are now increasingly using Large Language Models (LLMs) to continuously scan software updates to spot newly fixed bugs, and build attacks before fixes can be applied – all without human involvement.
The report titled ‘GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold’ from the Google Threat Intelligence Group noted a remarkable shift towards an autonomous offensive framework for threat actors. It also flagged a trend of “LLMJacking”, which involves reselling unauthorised access to high-performance models via underground proxy services, and threat actors using LLMs for target profiling as well as quickly exploring possible backdoors or weak spots to attack.
For the most part, attackers aren’t using new groundbreaking methods to break into systems – they are still going after known, unfixed issues, just with much greater speed and scale. Instead of relying on manual effort to keep up, we will have to embrace agentic defence. Or we risk falling behind,” said John Hultquist, Chief Analyst, Google Threat Intelligence Group.
Google Threat Intelligence Group and cybersecurity firm Mandiant said the threat actors are using or testing offensive AI frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI. These systems coordinate specialised agents to scan targets, discover weaknesses, generate attack payloads and adjust tactics based on the results.
A key evolution defining the speed and scale of cyberattacks is that, instead of a human hacker directing every action, AI frameworks can link several steps into an automated workflow.
One of the report’s most concerning findings regarding cybersecurity defence is the automation of vulnerability research. AI agents can now compare software code before and after a security patch or an update, identify the weakness that was supposedly fixed, generate an exploit and test it. That process could go either way, depending on the quality of the fix. This makes already disclosed vulnerabilities, known as n-days, into targets
Researchers have also identified a framework called exploit_agent that automated vulnerability analysis and exploitation without a human in the loop. It was associated with activity targeting websites in France, particularly the media, entertainment and education industries.
A key takeaway is that if security patches themselves become valuable intelligence for threat actors, any cybersecurity defence strategy now has considerably less time to apply further fixes.
The report references recent attacks targeting cryptocurrency platforms in the US and Europe, where the threat actor, named North Korea’s MIDNIGHT NEPTUNE, exploited LLM software architectures. They modified the .claude/settings.json configuration files, so that when users started the Claude Code, it automatically triggered a download for the backdoor called SOMBERMEME.
“To maintain stealth, suspected MIDNIGHT NEPTUNE further masked its lateral movement by generating GitHub Personal Access Tokens (PATs) named after Microsoft Copilot, underscoring how threat actors are simultaneously abusing generative tools for operational scaling and exploiting native AI agent execution hooks as enterprise access vectors,” the report explained.
Google made it clear that the defence against machine-speed threats at human speed isn’t feasible, and that some help in rethinking security in the agentic era could come from their Gemini and other frontier models, among other avenues.
“In addition to our proactive platform defences, we’ve recently introduced Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long-horizon capabilities. Google has built Gemini 4 Carbon with layered defences against cyber misuse – combining safety training, input/output classifiers, activation-based introspection probes and chain-of-thought and agentic trajectory monitors – designed to block requests to develop functional, weaponised exploits or malware tooling,” the researchers said.
Google rolled out Gemini 4 Carbon late last month for early testing, making it available only to AI Ultra subscribers this month. Carbon succeeded its predecessor frontier model, Gemini 4 Argon, in a matter of weeks, underlining the pace of evolution for frontier AI models.