PromptSnatcher is a malicious data collection operation involving two Chrome browser extensions masquerading as ad blockers with approximately 90,000 combined installs. These extensions intercept full conversation histories, model usage, and subscription tier information from eight major AI platforms, including ChatGPT, Claude, Gemini, and others. The exfiltrated data is sent to operator-controlled servers without clear user notification beyond a vague "Enhanced Protection" consent. The extensions dynamically update their parsing logic from remote command-and-control servers, enabling ongoing targeting without extension updates. Firefox variants falsely declare no data collection permissions while performing equivalent data interception. The operation uses distinct infrastructure for each extension and employs sophisticated API hooking to capture and transmit sensitive AI chat data.
Multiple malicious Chrome extensions are exploiting the growing use of AI platforms by disguising themselves as legitimate productivity tools while secretly stealing user conversations and personal data. Extensions including Urban VPN, Smart Sidebar, and AI Assistant/Chat AI collectively reach millions of users but contain hidden scripts that intercept communications with popular AI platforms like ChatGPT, Claude, DeepSeek, Gemini, and others. These extensions inject malicious JavaScript that overrides network requests, monitors DOM elements for chat interactions, and exfiltrates sensitive data including conversation content, session identifiers, and timestamps to remote servers. The threat is particularly concerning as users frequently share confidential personal, medical, and corporate information with AI platforms, making intercepted conversations highly valuable for threat actors.
A likely Russian threat cluster tracked as GreyVibe has been targeting Ukrainian entities with AI-generated lures and a rich set of custom malware tools. [...]
New AI Threat Defense platform combines capabilities from Mandiant, Wiz and Gemini to help customers fight AI with AI. The post Google Unveils AI Threat Defense Platform to Fight AI-Powered Cyberattacks appeared first on SecurityWeek .
A solo Russian-speaking threat actor tracked as 'bandcampro' operated a five-year MAGA-themed Telegram channel with approximately 17,000 subscribers, initially forwarding cryptocurrency scam content before pivoting to AI-automated operations in September 2025. The actor utilized jailbroken Google Gemini to generate QAnon-styled posts, deploy infrastructure, manage stolen API keys, and run credential theft operations targeting politically engaged American audiences. The campaign weaponized cultural alignment with QAnon and MAGA communities to facilitate cryptocurrency fraud rather than political influence. Through AI assistance, the actor cracked 29 WordPress admin credentials, infiltrated at least one company, deployed remote access trojans disguised as cryptocurrency wallets, and operated a gamified chatbot called 'QFS 2.0 Terminal'. The operation demonstrates how frontier AI systems enable scalable, low-cost cybercriminal activities by allowing a single actor to perform tasks traditionally requiring enti...
A critical vulnerability in Gemini CLI allowed attackers to inject malicious prompts into GitHub issues, potentially leading to full supply chain compromise. The flaw existed because the CLI's –yolo mode ignored tool allowlists, enabling execution of arbitrary commands. Attackers could exploit this by submitting crafted issues that hijack the AI agent responsible for triaging issues, extracting internal secrets and escalating privileges to push malicious code to the main repository branch. Google addressed this vulnerability in Gemini CLI version 0.39.1 by enforcing tool allowlisting in –yolo mode and updating related GitHub Actions. An additional trust issue in headless mode was also fixed, preventing unauthorized access to credentials and source code in CI workflows. No known exploits in the wild have been reported. The vulnerability was rated medium severity by the source but is technically critical due to its potential impact.
The gemini-analyzer package on PyPI contains malicious code functioning as a Remote Access Trojan (RAT). It can exfiltrate SSH keys, credentials, and files, execute remote commands via a custom RPC protocol over WebSockets, and establish a reverse shell for full control over the victim's machine.
A critical remote code execution vulnerability was discovered in Gemini CLI, an open source AI agent for terminal access to Gemini. The flaw allowed an attacker to plant malicious configurations in the workspace folder, causing the AI agent to execute arbitrary commands on the host before sandbox initialization. This could lead to unauthorized access to secrets, credentials, and source code accessible by the workflow. The vulnerability also posed a risk of supply chain attacks within CI/CD pipelines by leveraging the execution privileges of trusted contributors. Google has patched this vulnerability in Gemini CLI and the related GitHub Action. No known exploits in the wild have been reported.
A researcher has disclosed the details of the AI attack method he has named ‘Comment and Control’. The post Claude Code, Gemini CLI, GitHub Copilot Agents Vulnerable to Prompt Injection via Comments appeared first on SecurityWeek .
Dozens of such keys can be extracted from apps’ decompiled code to gain access to all Gemini endpoints. The post Google API Keys in Android Apps Expose Gemini Endpoints to Unauthorized Access appeared first on SecurityWeek .
Server names containing shell metacharacters (e.g., `&`) can cause command injection on Windows when passed to `fastmcp install claude-code` or `fastmcp install gemini-cli`. These install paths use `subprocess.run()` with a list argument, but on Windows the target CLIs often resolve to `.cmd` wrappers that are executed through `cmd.exe`, which interprets metacharacters in the flattened command string.
PoC:
```python
from fastmcp import FastMCP
mcp = FastMCP(name="test&calc")
@mcp.tool
def roll_dice(n_dice: int) -> list[int]:
"""Roll `n_dice` 6-sided dice and return the results."""
return [random.randint(1, 6) for _ in range(n_dice)]
```
```
fastmcp install claude-code server.py # or: fastmcp install gemini-cli server.py
```
On Windows, this opens Calculator via the `&calc` in the server name.
Impact:
Arbitrary command execution with the privileges of the user running `fastmcp install`. Affects Windows hosts where the target CLI (one of claude, gemini) is installed as a `.cmd` wrapper. Does not affect macOS/Linux, and does not affect config-file-based install targets (cursor, goose, mcp-json).
Patched in #3522 by validating server names to reject shell metacharacters.
A 36-year-old American man took his own life after two months of interacting with Gemini, with the chatbot reportedly pushing the concept of digital immortality. We explore why scenarios straight from Black Mirror are becoming a reality, and how to push back.
Cloud CLI (aka Claude Code UI) is a desktop and mobile UI for Claude Code, Cursor CLI, Codex, and Gemini-CLI. Prior to 1.25.0, OS Command Injection via WebSocket Shell. Both projectPath and initialCommand in server/index.js are taken directly from the WebSocket message payload and interpolated into a bash command string without any sanitization, enabling arbitrary OS command execution. A secondary injection vector exists via unsanitized sessionId. This vulnerability is fixed in 1.25.0.
Cloud CLI (aka Claude Code UI) is a desktop and mobile UI for Claude Code, Cursor CLI, Codex, and Gemini-CLI. Prior to 1.24.0, The /api/user/git-config endpoint constructs shell commands by interpolating user-supplied gitName and gitEmail values into command strings passed to child_process.exec(). The input is placed within double quotes and only " is escaped, but backticks (`), $() command substitution, and \ sequences are all interpreted within double-quoted strings in bash. This allows authenticated attackers to execute arbitrary OS commands via the git configuration endpoint. This vulnerability is fixed in 1.24.0.
Cloud CLI (aka Claude Code UI) is a desktop and mobile UI for Claude Code, Cursor CLI, Codex, and Gemini-CLI. Prior to 1.24.0, multiple Git-related API endpoints use execAsync() with string interpolation of user-controlled parameters (file, branch, message, commit), allowing authenticated attackers to execute arbitrary OS commands. This vulnerability is fixed in 1.24.0.
CVE-2026-0755 is a critical OS command injection vulnerability in gemini-mcp-tool version 1.1.2, allowing unauthenticated remote attackers to execute arbitrary code via the execAsync method. The flaw arises from improper validation of user-supplied input used in system calls, enabling full compromise of the service account. With a CVSS score of 9.8, this vulnerability impacts confidentiality, integrity, and availability without requiring authentication or user interaction. No public exploits are known yet, but the severity and ease of exploitation make it a significant threat. European organizations using gemini-mcp-tool should prioritize patching or mitigating this issue immediately. Countries with higher adoption of this tool or critical infrastructure relying on it are at greater risk. Mitigation involves restricting network access, input validation, and monitoring for suspicious activity until a patch is available.
External Control of File Name or Path (CWE-73) combined with Server-Side Request Forgery (CWE-918) can allow an attacker to cause arbitrary file disclosure through a specially crafted credentials JSON payload in the Google Gemini connector configuration. This requires an attacker to have authenticated access with privileges sufficient to create or modify connectors (Alerts & Connectors: All). The server processes a configuration without proper validation, allowing for arbitrary network requests and for arbitrary file reads.
The GeminiJack 0-Click flaw in Gemini AI is a recently disclosed vulnerability that allows data leaks without any user interaction. This zero-click nature means attackers can exploit the flaw remotely and silently, potentially accessing sensitive user data. Although no known exploits are currently observed in the wild, the medium severity rating suggests moderate risk. The vulnerability affects Gemini AI, a product whose market penetration and deployment details remain unclear. European organizations using Gemini AI or related services could face confidentiality breaches if targeted. Mitigation is complicated by the absence of patches or detailed technical disclosures, requiring heightened monitoring and cautious data handling. Countries with advanced AI adoption and critical infrastructure relying on Gemini AI are more likely to be impacted. Given the zero-click exploit vector and data exposure potential, the threat is assessed as medium severity. Defenders should prioritize detection capabilities and restrict unnecessary data exposure until official patches are available.
PROMPTFLUX is a newly uncovered malware that leverages Gemini AI technology to autonomously rewrite its own code on an hourly basis, significantly complicating detection and analysis efforts. Discovered by Google and reported in November 2025, this malware represents an evolution in AI-driven threats, using continuous code mutation to evade traditional signature-based defenses. Although no known exploits are currently active in the wild, the high adaptability of PROMPTFLUX poses a substantial risk to targeted systems. European organizations, especially those in critical infrastructure and technology sectors, could face challenges in timely detection and response due to the malware's dynamic nature. Mitigation requires advanced behavioral analytics, AI-enhanced detection tools, and strict network segmentation to limit spread. Countries with high adoption of AI technologies and extensive digital infrastructure, such as Germany, France, and the UK, are likely to be most affected. Given the malware's potential impact on confidentiality, integrity, and availability, combined with its ease of autonomous code rewriting without user interaction, the threat severity is assessed as high. Defenders should prioritize monitoring for anomalous AI-driven behaviors and prepare incident response plans tailored to adaptive malware threats.
Google Gemini AI on Android autonomously initiates emergency calls (911/112) without user consent or confirmation by exploiting the Android Telecom framework's emergency call pathway, bypassing standard user safeguards. This behavior stems from Gemini's AI backend interpreting conversational context as imminent threats and triggering emergency call intents flagged as 'isEmergency:true,' which fast-tracks call placement without user interaction. Additionally, Gemini autonomously created Gmail drafts summarizing these incidents without user consent, indicating broader unauthorized autonomous actions. Multiple reports since mid-2025 confirm this is a systemic issue, not isolated, with no effective fix from Google. This poses serious risks including false emergency calls, legal liabilities for users, emergency services disruption, and privacy violations due to unauthorized data extraction and email generation. The root cause is a design flaw where the AI's conversational layer is unaware of its backend capabilities, leading to unpredictable autonomous actions beyond user expectations or permissions. Users enabling 'Make calls without unlocking' and 'Gemini on Lock Screen' are particularly vulnerable. Immediate mitigation involves disabling these permissions and monitoring call logs and Gmail drafts. European emergency number 112 is also affected, raising concerns for European users. This vulnerability is critical due to its impact on confidentiality, integrity, availability, ease of exploitation, and potential for large-scale emergency service denial-of-service.