Google Threat Intelligence Group has published a report detailing how a functional AI exploit has transitioned artificial intelligence from a digital assistant to an active participant in cyberattacks. For the first time, researchers recorded an instance where AI developed a functional exploit for mass attacks almost from scratch.
NEWSROOM IN examines the severity of this development and whether current security measures remain effective.
AI Exploit Potential: What Has Changed for Bad Actors?
Previously, hackers primarily used chatbots to craft convincing phishing emails. Now, state-sponsored groups from China and North Korea, such as the notorious APT45, are utilizing neural networks as “security auditors.”
- Finding Zero-Day Vulnerabilities: These are the most dangerous security flaws that are unknown even to developers. AI has learned to identify not just technical errors, but logical contradictions in code that standard scanners typically miss.
- Hallucinations as Evidence: Google experts identified an AI-written exploit by specific markers, including training-related comments in the code, non-existent security ratings, and formatting typical of Python-based AI models.
The PROMPTSPY Backdoor: Remote Device Control
A significant example of this new era is the Android malware PROMPTSPY. This program uses the Gemini model to analyze a user’s screen by reading the interface structure. It sends this data to the neural network and receives commands on where to click or swipe.
PROMPTSPY can independently enter PIN codes, intercept pattern locks, and prevent its own deletion by creating invisible elements over “Uninstall” buttons. Effectively, a “ghost user” controls the device with greater speed and precision than the owner.
Noise Tactics: AI-Generated Junk Code
To evade antivirus detection, hackers use AI to create “informational noise.” Programs like CANFAIL and LONGSTREAM are filled with thousands of lines of useless AI-written code. For example, the software may include dozens of daylight savings time checks that perform no function but make the program appear legitimate to security scanners.
The Main Question: Is It Time to Worry?
The current verdict remains cautious: the situation is manageable for now, but with certain caveats.
Why panic is not yet necessary:
- Google is staying ahead: In the recent case, company experts detected the AI exploit before it could be deployed.
- Shield against the sword: Google uses the same AI technology to identify and automatically patch vulnerabilities, creating an algorithmic arms race where defensive models are equally powerful.
Reasons for concern:
- Autonomy: Hackers are moving toward systems like Hexstrike, which can conduct reconnaissance and coordinate attacks without human intervention.
- Supply chain trust: Attackers have begun targeting popular GitHub libraries, such as LiteLLM, which are used by AI developers worldwide.
Expert Advice: In the era of AI hackers, standard two-factor authentication (2FA) via SMS is becoming increasingly vulnerable. Users should switch to hardware security keys or authenticator apps. AI can analyze logic, but it cannot bypass a physical key.
