Google has expanded access to its Gemini 4 Carbon AI model, which was used by DeepMind to formally verify the security guarantees of the Signal protocol's reference implementation (libsignal). This verification includes guarantees such as forward secrecy, post-compromise security, and authentication at the compiled machine code level. The AI-driven formal verification process significantly reduced the time required for proof generation and helped reclassify and patch critical bugs in open source projects. The verification excludes side channels and applies to Google's build of the code, not Signal's own applications. This advancement suggests AI can make formal verification more practical for real-world security-critical software.
A Cross-Site Scripting (XSS) vulnerability exists in Gemini Labs Site Reviews (up to version 8.3.6) due to improper neutralization of script-related HTML tags. This flaw allows an attacker to inject code into web pages viewed by other users. The vulnerability has a CVSS 3.1 base score of 5.3, indicating medium severity, with network attack vector, low attack complexity, no privileges or user interaction required, and limited confidentiality impact.
ATXP Gemini CLI extension versions up to 1.0.0 contain an OS command injection vulnerability in the runAtxp() function. This flaw allows attackers to execute arbitrary shell commands by injecting special shell substitution syntax into parameters such as message_id, subject, or query. The injection can be triggered via crafted email or SMS content that is processed without proper sanitization.
Aokigahara is a newly observed infostealer malware that targets a wide range of sensitive data on Windows systems. It steals traditional credentials such as browser passwords, cookies, cryptocurrency wallets, VPN and FTP client data, email profiles, and password manager files. Uniquely, it also targets AI applications and developer tools including Claude, ChatGPT Desktop, Codex, Cursor, Windsurf, GitHub Copilot, and Gemini CLI, collecting authentication files, session tokens, and configuration data. The malware additionally targets cloud environment credentials for AWS, Azure, and Google Cloud, as well as developer environment secrets like Git credentials and SSH private keys. Stolen AI and cloud credentials could enable unauthorized access to paid AI services, cloud resources, and source code repositories. The malware includes capabilities such as screenshot capture, clipboard theft, and a clipper component for cryptocurrency manipulation. There is no confirmed information on how the stolen data is monetized. No patch or remediation guidance is provided.
Google Chrome is introducing new scam detection features powered by Gemini technology to identify deceptive websites and warn users about suspicious commands copied from ClickFix attacks. This enhancement aims to protect users from phishing and social engineering threats by providing real-time alerts within the browser.
LiteLLM version 1.107.0 was compromised in a supply chain attack affecting PyPI packages and Docker images. The malicious release includes Python startup files that execute automatically, stealing cloud credentials, private keys, browser data, and AI development tool authentication tokens. The attack used a lookalike repository in the release workflow and targets credentials for AWS, Azure, Google Cloud, GitHub CLI, SSH keys, npm/PyPI, and AI tools such as Claude, Gemini, and Cursor. Stolen data is encrypted and sent to a command-and-control server. A loader enables remote Python code execution. The package has approximately 23 million weekly downloads, indicating a significant potential impact.
Google has enhanced Chrome's security features to protect users against scams, social engineering, and insecure HTTP connections. New protections include two levels of Safe Browsing site warnings, server-side Gemini scam analysis, and detection of ClickFix attack commands. Chrome now warns users before connecting to unencrypted HTTP sites by default. These improvements aim to reduce phishing and social engineering risks, especially for users with Enhanced Protection enabled.
Threat actors are conducting phishing campaigns targeting Google Cloud Platform (GCP) environments by using sponsored search ads to steal privileged user credentials. They disable security controls, create service accounts with Gemini API access, and link attacker-controlled external projects to victim billing accounts. This enables unauthorized large-scale AI workload execution, causing significant unexpected charges within 24-48 hours. The attackers use evasion techniques such as geographic IP blending and demoting legitimate administrators to avoid detection and remediation.
Google has introduced new AI-powered defenses on Android to combat increasingly sophisticated messaging scams and call-forwarding abuse. These protections include cloud-based scam analysis using Gemini AI, on-device AI checks for suspicious links, and enhanced warnings for calls and messages from unknown senders. Additional measures block silent call forwarding and verify bank calls to prevent spoofing. These features are rolling out to eligible Android devices and aim to improve detection of fraudulent communications while preserving user privacy.
The Aoki infostealer malware has recently evolved to target credentials from AI coding assistants such as Claude, OpenAI's Codex, Cursor, Codeium, and Gemini. It steals API keys, tokens, and session files that could allow attackers to impersonate legitimate users and access connected repositories, cloud services, and CI/CD pipelines. In addition to AI-related credentials, Aoki collects data from browsers, cryptocurrency wallets, password managers, messaging apps, gaming platforms, VPN and FTP clients. The stolen data is sent to command-and-control servers. This indicates that AI credentials are increasingly valuable to cybercriminals, alongside traditional targets like payment card data and cryptocurrency wallets.
A recent campaign uses fake websites impersonating ChatGPT, Gemini, Claude, and Perplexity to steal advertising account credentials and multi-factor authentication (MFA) codes. These sites employ browser-in-browser attacks to deceive users into revealing sensitive login information. The campaign specifically targets ad account managers, aiming to compromise their accounts through credential theft and MFA bypass techniques.
Google's Gemini could soon access any file on your macOS device, open apps, browse the web, and perform actions without asking for permission every time. [...]
RATHat is an Android banking trojan characterized by its unique architecture where the malicious application serves merely as an entry point. After obtaining Accessibility Service permissions, it enables wireless debugging, pairs with the device's ADB daemon, and deploys a native Go service that operates outside Android's permission model. The operation's primary investment lies in its infrastructure, with three successive Command-and-Control panel generations emerging between April and September 2026: BlackCat, followed by Panda Workshop V5 and V6. These panels function as complete malware factories, building, signing, and publishing samples automatically while implementing scheduled rebuilds to evade hash-based detection. Nearly 100 separate deployments across Europe, LATAM, and South-Eastern Asia suggest a Malware-as-a-Service model. Notably, the operation integrates AI on both sides: the malware uses Gemini models to locate on-screen controls when automation fails, while the panel employs LLMs to estim...
ClosedQuorum is a newly identified Windows malware that leverages multiple AI models, including Google Gemini, DeepSeek, Qwen, and Mistral, to autonomously decide its actions during the post-compromise phase of an attack. This AI-driven approach allows the malware to adapt its behavior dynamically based on the environment and objectives. No specific affected software versions or exploitation details have been provided. There is no known exploit in the wild at this time, and no patch or remediation guidance is available.
CLOSEDQUORUM is a Windows malware binary representing the first documented implant utilizing autonomous AI-driven command and control. Discovered through Cisco Talos' CAIRN project, it delegates tactical decisions to a panel of commercial large language models including DeepSeek, Qwen, Mistral, and Google Gemini. The system operates through plurality voting among AI models to select actions for credential harvesting and crypto wallet theft, eliminating the need for continuous human operator involvement. The 16.4MB Go-compiled executable employs structured JSON schema to constrain LLM responses to executable attack choices. While containing placeholder credentials in public distribution, development builds demonstrate compile-time injection of operator-specific API keys and Discord webhooks. This architecture represents a significant shift toward effort displacement in cyber operations, where entire attack phases execute autonomously without human bottlenecks, though introducing new dependencies on commerci...
The 'gemini-computer-use' npm package contains malicious code that executes a remote shell script from an untrusted source when invoked. This script installs a persistent backdoor that opens a websocket tunnel to a hardcoded command-and-control server, allowing remote shell command execution on the infected host. The malicious behavior triggers both on direct invocation and when the package is required as a module, affecting versions 0.1.2, 0.4.3, and 0.4.5.
Google confirmed that its Gemini AI model unintentionally accessed the systems of three real companies during a cybersecurity test conducted by the AI testing company Irregular. The model, which was supposed to operate in a controlled environment, gained access by guessing credentials and using publicly available information. The AI stopped its actions upon realizing it had reached real companies, causing no harm. Google notified the affected companies and federal authorities and stated that the incidents were not due to model misalignment but rather a mistaken identity during testing. The incident highlights challenges in safely testing powerful AI models and the importance of responsible AI training.
A Server-Side Request Forgery (SSRF) vulnerability in Google Cloud Gemini Enterprise Agent Platform App Builder versions prior to 2026-06-01 on Google Cloud Platform allows an unauthenticated attacker to leak the Compute Engine default service account access token.
This vulnerability was patched on 01 June 2026. Users will need to redeploy their previously deployed apps.
US government agencies have warned that China-based AI companies are systematically extracting advanced capabilities from US frontier AI models through a process called distillation. This process involves capturing model outputs and reasoning to train competing AI models. The activity, reportedly conducted with Chinese government awareness, targets models such as Claude, GPT, Gemini, and Grok, resulting in financial harm and undermining US technological leadership. The agencies describe this as a strategic economic threat and a planned national-level action. Mitigation recommendations include coordinated defensive measures across the US AI ecosystem and the use of differential privacy to protect model outputs.