Google Cloud

GTIG AI Threat Tracker: From Prompting to Autonomy  The Evolution of Adversarial AI

Chinese hackers are running AI on stolen networks to avoid detection, Google says

The Chosun Daily

AI Cyberattacks Steal Thousands in 6 Hours

Gizmodo

Silicon Valley’s AI Agent Push Has Been Paying Off—for Cybercriminals

Belfer Center

Deception and Detection: Why Artificial Intelligence Empowers Cyber Defense over Offense

Openai

The Defenders Window

Infosecurity Magazine

Agentic AI will Supercharge Cyber Threats. But Not the Way You Thin

Center for a New American Security

Tipping the Scales

US Department of War

Csa_China_Based_Ai_Companies_Malicious_Distillation_Against_Us.Pdf

Metaculus

Will cyberattack targeting AI cause power blackout in the US before 2028?

Google's Threat Intelligence Group (GITG) warned that hackers and other cybercriminals are using AI in ever more sophisticated ways as part of their activities, which are increasingly targeting AI assets, in a report published Tuesday.

Google's Threat Intelligence Group (GITG) warned that hackers and other cybercriminals are using AI in ever more sophisticated ways as part of their activities, which are increasingly targeting AI assets, in a report published Tuesday.

In one instance, GITG observed an unspecified threat actor use an AI coding chatbot, a prompt and a series of agentic instructions to plot, develop and execute a mass harvesting campaign that compromised thousands of third-party credentials in less than six hours.

In one instance, GITG observed an unspecified threat actor use an AI coding chatbot, a prompt and a series of agentic instructions to plot, develop and execute a mass harvesting campaign that compromised thousands of third-party credentials in less than six hours.

The unit also tracked UNC6508 — a Chinese outfit targeting North American medical, academic and military research institutions since 2023 — deploying open-source AI models inside compromised cloud environments to use their resources while avoiding detection by commercial APIs.

The unit also tracked UNC6508 — a Chinese outfit targeting North American medical, academic and military research institutions since 2023 — deploying open-source AI models inside compromised cloud environments to use their resources while avoiding detection by commercial APIs.

Separately, the report said a group labelled UNC6780, also known as TeamPCP, embedded prompts referencing biological and nuclear weapons in malware loader scripts to make AI security scanners refuse or skip analysis of the malicious code beneath, among other payload delivery methods.

Separately, the report said a group labelled UNC6780, also known as TeamPCP, embedded prompts referencing biological and nuclear weapons in malware loader scripts to make AI security scanners refuse or skip analysis of the malicious code beneath, among other payload delivery methods.

Cyber actors across China, Russia and Iran, meanwhile, were found leveraging AI tools such as Gemini to increase the productivity of their influence operations, including using highly specific prompts to create "photorealistic" media content and refine narratives.

Cyber actors across China, Russia and Iran, meanwhile, were found leveraging AI tools such as Gemini to increase the productivity of their influence operations, including using highly specific prompts to create "photorealistic" media content and refine narratives.

The findings were released the same day that the NSA, CISA and the FBI published a joint advisory accusing six China-based AI companies, including DeepSeek, Moonshot AI and Alibaba, of extracting billions of tokens from U.S. frontier models such as Claude, GPT, Gemini and Grok.

Donald Trump

The findings were released the same day that the NSA, CISA and the FBI published a joint advisory accusing six China-based AI companies, including DeepSeek, Moonshot AI and Alibaba, of extracting billions of tokens from U.S. frontier models such as Claude, GPT, Gemini and Grok.

Google Cloud #

NBC #3261

The Chosun Daily #

Gizmodo #

Belfer Center #

Openai #

Infosecurity Magazine #

Center for a New American Security #

US Department of War #

Metaculus #

Google's Threat Intelligence Group said Tuesday that hackers have shifted from basic AI prompting to agentic workflows, in one case compromising a cloud resource and then planning, building and executing a mass credential harvesting campaign in under six hours.

Google's Threat Intelligence Group said Tuesday that hackers have shifted from basic AI prompting to agentic workflows, in one case compromising a cloud resource and then planning, building and executing a mass credential harvesting campaign in under six hours.

Google said a group it tracks as UNC6508, linked to China, has targeted academic, medical and military research institutions in North America since 2023, focusing on proprietary AI research. Google did not identify any of the victims.

Google said a group it tracks as UNC6508, linked to China, has targeted academic, medical and military research institutions in North America since 2023, focusing on proprietary AI research. Google did not identify any of the victims.

According to the report, UNC6508 deployed open-source AI models inside compromised cloud environments, which lets attackers use victims' computing resources while avoiding the monitoring built into commercial AI application programming interfaces.

Ynetglobal

According to the report, UNC6508 deployed open-source AI models inside compromised cloud environments, which lets attackers use victims' computing resources while avoiding the monitoring built into commercial AI application programming interfaces.

Liu Chang, a spokesperson for the Chinese Embassy in Washington, rejected the allegations, saying China "opposes hacking activities and fights such activities in accordance with the law" while rejecting what he called smears under the pretext of cybersecurity.

Liu Chang, a spokesperson for the Chinese Embassy in Washington, rejected the allegations, saying China "opposes hacking activities and fights such activities in accordance with the law" while rejecting what he called smears under the pretext of cybersecurity.

The report said a group tracked as UNC6780, also called TeamPCP, embedded prompts referencing biological and nuclear weapons in malware loader scripts, aiming to make AI security scanners refuse or skip analysis of the malicious code beneath.

The report said a group tracked as UNC6780, also called TeamPCP, embedded prompts referencing biological and nuclear weapons in malware loader scripts, aiming to make AI security scanners refuse or skip analysis of the malicious code beneath.

The National Security Agency, CISA and the FBI published a joint advisory Sept. 8 accusing six China-based AI firms, including DeepSeek, Moonshot AI and Alibaba, of extracting billions of tokens from U.S. frontier models such as Claude, GPT, Gemini and Grok.

International Business Times UK

The National Security Agency, CISA and the FBI published a joint advisory Sept. 8 accusing six China-based AI firms, including DeepSeek, Moonshot AI and Alibaba, of extracting billions of tokens from U.S. frontier models such as Claude, GPT, Gemini and Grok.

AI panic in cybersecurity is overblown, because machines are far better at spotting patterns than at the creative deception offense demands. Defenders can now hand an agent a codebase and watch it find and patch flaws that humans never had the time or expertise to reach. Invest in defensive automation now and the internet ends up more secure than it has ever been.

AI panic in cybersecurity is overblown, because machines are far better at spotting patterns than at the creative deception offense demands. Defenders can now hand an agent a codebase and watch it find and patch flaws that humans never had the time or expertise to reach. Invest in defensive automation now and the internet ends up more secure than it has ever been.

Attackers only have to win once, while defenders have to guard everything, and AI widens that gap instead of closing it. Agentic systems probe the fence continuously at machine speed, adapting the instant one route gets blocked, turning occasional attacks into constant pressure. Well-resourced attackers with a high risk appetite can tolerate failures and collateral damage that defenders never could.

Attackers only have to win once, while defenders have to guard everything, and AI widens that gap instead of closing it. Agentic systems probe the fence continuously at machine speed, adapting the instant one route gets blocked, turning occasional attacks into constant pressure. Well-resourced attackers with a high risk appetite can tolerate failures and collateral damage that defenders never could.

Artificial Intelligence

Google: Hackers Automate Attacks With Agentic AI

Google: Hackers Automate Attacks With Agentic AI

United States of America

Cybercrime

Google's findings conclusively prove that AI is widening the gap between defenders and attackers in the latter's favor. Agentic systems can probe the fence continuously at machine speed to find weaknesses, which, when they do, they exploit ruthlessly. When backed by the resources of hostile states, the threat intensifies even further, to everyone's detriment.

Google's findings conclusively prove that AI is widening the gap between defenders and attackers in the latter's favor. Agentic systems can probe the fence continuously at machine speed to find weaknesses, which, when they do, they exploit ruthlessly. When backed by the resources of hostile states, the threat intensifies even further, to everyone's detriment.

While cybercriminals and hackers will use AI as a weapon, the advantage lies squarely with the defenders. Using AI as part of their arsenal, defenders can now use agents to find and patch flaws that humans never had the time or expertise to identify. In this way, defensive automation promises to build a near-impenetrable cyber wall to block the machinations of hostile actors.

While cybercriminals and hackers will use AI as a weapon, the advantage lies squarely with the defenders. Using AI as part of their arsenal, defenders can now use agents to find and patch flaws that humans never had the time or expertise to identify. In this way, defensive automation promises to build a near-impenetrable cyber wall to block the machinations of hostile actors.

There is a 10% chance that a cyberattack targeting AI systems will cause a significant power blackout in the United States before 2028, according to the Metaculus prediction community.

There is a 10% chance that a cyberattack targeting AI systems will cause a significant power blackout in the United States before 2028, according to the Metaculus prediction community.

Google: Hackers Now Use AI Agents to Attack in Hours

Google: Hackers AutomateNow AttacksUse WithAI AgenticAgents AIto Attack in Hours

The Spin

Metaculus Prediction

Go Deeper