prompt injection

AppWizard
July 23, 2026
GitHub will reject command-line support bundle uploads from outdated versions of GitHub Enterprise Server lacking security patches starting August 18, 2026. The npm package @copilot-mcp/apex has been identified as a post-install dropper that installs a macOS infostealer, phishing for sensitive information and maintaining a connection to an attacker's server. A rogue extension on the Microsoft Visual Studio Code marketplace, "Markdown All Pro," impersonates a legitimate tool and opens a backdoor after installation. A phishing campaign targeting Portuguese users delivers the Lampion banking malware, which has been active since 2019. DoubleVerify reports a rise in "AfterCall" apps that exploit user permissions for ad fraud. The GhostCommit attack method hides malicious instructions within PNG images in pull requests. The U.S. government has updated its advisory on Iranian-affiliated cyber activity targeting operational technology devices. An Android app posing as a civil defense alert system has been found to contain malware for data harvesting. An Iranian threat actor is distributing MarkiRAT malware through fake applications. An analysis of 28 AI-coded applications revealed 434 vulnerabilities, prompting Cisco to introduce Antares to identify vulnerabilities in codebases. A Russian-speaking threat actor is dismantling guardrails on AI models to create offensive tools.
Winsage
June 19, 2026
Microsoft has introduced the Microsoft Execution Containers (MXC) SDK to establish Windows as a reliable operating system for autonomous agents, focusing on containment, identity, and manageability. The MXC framework serves as a policy-driven execution layer for agents on Windows and Windows Subsystem for Linux (WSL), allowing developers to set access permissions using JSON or TypeScript. It employs process and session isolation for agent containment and identity. Future enhancements will include micro-VM support for high-risk tasks and integration with Windows 365 for cloud PC workloads. IT teams can manage MXC policies through Entra ID and Intune, while Defender and Purview provide protection and observability. The MXC framework is built on Microsoft's security initiatives, including Secure Boot and passwordless sign-in, allowing agents to inherit a secure foundation. However, early commentary expresses caution regarding MXC's perception as a comprehensive security solution, noting issues with overly permissive policies and the lack of outbound network filtering. Other platforms, such as Linux, are also enhancing security for agents with kernel-level isolation and secure environments like NVIDIA's OpenShell runtime. Various projects are focusing on agent sandboxes within Kubernetes, employing technologies like gVisor and Kata Containers for isolation. Overall, no singular dominant platform security model for AI agents has emerged, with Windows' MXC still considered nascent compared to existing solutions in Linux and Kubernetes ecosystems.
Winsage
June 9, 2026
Microsoft announced a suite of agent-centric features for Windows 11 at the Build 2026 conference, introducing the Microsoft Agent Platform for local AI agents. This includes enhanced integration with Microsoft Foundry, GitHub, and the M365 suite. The developer documentation covers local agent runtimes and isolation patterns, highlighting the Microsoft Execution Containers SDK and Entra Agent ID. Enterprise controls like Intune and Agent 365 are also featured. The upcoming Copilot Actions will allow agents to interact with local files and applications, aimed at improving workflows. Microsoft is addressing security concerns with resources outlining strategies to mitigate risks such as cross-prompt injection. Industry commentary indicates a growing discussion about privacy and security in relation to these changes.
Winsage
June 2, 2026
AI agents have evolved from simple question-answering systems to autonomous entities that can perform actions across various platforms. This shift raises concerns about control and trust, necessitating a change in security paradigms. Developers are now required to integrate security into the architecture of their platforms to maintain trust in agent deployment. Microsoft has expanded Agent 365 to manage local agents on Windows, introducing policy-based controls to govern agent actions. The Microsoft Execution Containers (MXC) SDK provides a policy-driven execution layer for agents, allowing developers to define constraints and ensuring consistent enforcement at runtime. Windows supports various containment options, including process and session isolation, to mitigate risks associated with agent behavior. Micro-VMs and Linux containers are also being integrated into the containment model. Windows 365 for Agents enables agents to operate in a managed cloud environment, limiting potential compromises. Collaborations with industry leaders aim to align containment strategies with developer needs. The security model is built on a foundation designed to minimize risk, incorporating features like passwordless sign-in and real-time protection through Windows Defender. The focus remains on enabling developers to create secure, governable agents for real-world deployment.
AppWizard
May 12, 2026
Google will integrate its Gemini 3.1 AI into the Chrome toolbar for Android starting in June, allowing users to summarize articles, ask questions about content, and extract details without switching apps. Users can enable the "Personal Intelligence" feature for tailored responses based on personal preferences. The Nano Banana feature will let users create or modify visuals from web pages. The auto browse function will allow Chrome to perform tasks like reserving parking or updating orders automatically. These features will include security protections, but sensitive actions will still require user confirmation. Gemini in Chrome will require devices with at least 4GB of RAM, running Android 12 or newer, and set to English-U.S. The rollout will begin for select Android devices in the U.S. at the end of June, with the auto browse feature available initially only to AI Pro and Ultra subscribers.
Winsage
April 5, 2026
On April 3, 2026, Anthropic expanded Claude’s desktop control feature to Windows for Pro and Max subscribers, allowing users to operate applications, navigate web pages, and manage files on their PCs without prior configuration. The feature is in research preview and includes a Dispatch companion for task assignment from mobile devices. Claude uses a structured tool hierarchy for task execution, prioritizing connectors like Slack and Google Calendar, and engages in direct desktop control only when necessary. Users must opt in to activate the feature, which integrates with existing software without requiring API keys. The technology is partly derived from Anthropic’s acquisition of Vercept AI, which specializes in AI-driven computer control. Security concerns have arisen due to vulnerabilities demonstrated shortly after the launch, prompting Anthropic to implement safeguards while acknowledging the feature's potential errors. Users can stop Claude's operations, but the company admits it cannot disable the technology remotely once tasks have started. Competitors like Microsoft and Google are also exploring similar desktop-level AI automation capabilities.
Winsage
February 11, 2026
Microsoft has released updates addressing over 50 vulnerabilities in its Windows operating systems and applications, including six critical zero-day vulnerabilities. 1. CVE-2026-21510: A security feature bypass in Windows Shell that allows execution of malicious content via a single click on a link, affecting all supported Windows versions. 2. CVE-2026-21513: Targets MSHTML, the web browser engine in Windows. 3. CVE-2026-21514: A security feature bypass in Microsoft Word. 4. CVE-2026-21533: Allows local attackers to gain SYSTEM level access in Windows Remote Desktop Services. 5. CVE-2026-21519: An elevation of privilege flaw in the Desktop Window Manager (DWM). 6. CVE-2026-21525: A potential denial-of-service threat in the Windows Remote Access Connection Manager. Additionally, the updates include fixes for remote code execution vulnerabilities affecting GitHub Copilot and various IDEs, specifically CVE-2026-21516, CVE-2026-21523, and CVE-2026-21256, which arise from a command injection flaw. Security experts emphasize the importance of safeguarding developers due to their access to sensitive data and recommend applying least-privilege principles.
AppWizard
February 10, 2026
Google Translate’s new AI-powered Advanced mode can engage in conversation rather than just translating text due to "prompt injection," which causes the model to struggle with distinguishing between translation requests and instructions. Users have found that this mode, based on a Gemini-based large language model, can respond to inquiries rather than providing straightforward translations. The older Classic mode remains a reliable option for consistent translations without unexpected interactions.
Winsage
December 11, 2025
Microsoft's December Patch Tuesday update addresses three critical zero-day vulnerabilities and a total of 56 bugs, including: - 28 elevation-of-privilege vulnerabilities - 19 remote-code-execution vulnerabilities - 4 information-disclosure vulnerabilities - 3 denial-of-service vulnerabilities - 2 spoofing vulnerabilities Three remote code execution flaws are classified as "critical." One zero-day vulnerability, CVE-2025-62221, allows attackers to gain SYSTEM privileges through the Windows Cloud Files Mini Filter Driver. The other two vulnerabilities fixed are: - CVE-2025-64671: A remote code execution vulnerability in GitHub Copilot for Jetbrains, exploitable via Cross Prompt Injection. - CVE-2025-54100: A PowerShell remote code execution vulnerability that can execute scripts from a webpage using Invoke-WebRequest. CVE-2025-62221 is attributed to MSTIC and MSRC, CVE-2025-64671 was disclosed by Ari Marzuk, and CVE-2025-54100 was identified by multiple security researchers.
Search