AI Sec Watch

Shadow AI (unauthorized AI agents built within organizations) has shifted from a data leakage risk to an access control problem. Unlike passive tools where employees paste data into public AI services, AI agents are active systems that can call APIs (application programming interfaces, which let software talk to other software), use stored credentials, and take actions in production systems without human approval for each step. Existing security controls designed for human users don't detect or manage these agents, which accumulate broad permissions and remain active even after employees leave.

The Advisory Forum is a governance body established under the EU AI Act to provide technical expertise and advice to the European Commission and AI Board on implementing the Act. It consists of 174 members representing balanced stakeholder groups (industry, startups, SMEs, civil society, and academia) plus five permanent member organizations, and was officially appointed on June 1, 2026.

Security operations centers (SOCs, teams that monitor and respond to security threats) have traditionally faced unavoidable trade-offs between three goals: quality (thorough investigation), consistency (standardized processes), and cost efficiency. This constraint exists because SOCs rely on human analysts to triage, investigate, and resolve alerts, which limits how much of each goal can be achieved simultaneously. Modern SOCs are hitting the limits of this model as alert volumes grow and work becomes more complex, forcing organizations to choose between degraded quality, inconsistent decisions, or higher costs.

This guide advises security leaders at small and medium-sized businesses (SMBs) on safely adopting Claude AI tools by understanding which Claude plan and products (Code, Cowork, Chat) match business needs, using a phased approval process to control risk exposure, and gradually enabling features rather than all at once. The text emphasizes that the AI landscape changes rapidly, shadow AI use (employees using unlicensed AI tools) is widespread, and security teams should risk-rank Claude's features before enabling them, being cautious about features like web search and browser extensions that could enable indirect prompt injection (attacks hidden in external content that trick the AI into following unintended instructions).

Microsoft discovered a security vulnerability called "AutoJack" that allows malicious webpages to trick AI agents (programs that can browse the web and access local services) into running harmful code on a user's computer. The attack works by chaining together three separate weaknesses in AutoGen Studio (Microsoft's tool for building AI agents), exploiting the fact that web-browsing agents have trusted access to local services that normally block outside access.

Cisco announced it will acquire WideField Security, a company that specializes in identity lifecycle security (managing who can access systems and what they can do), to enhance Splunk's Agentic SOC (a security operations center that uses AI agents to automate threat detection). WideField's technology helps organizations discover identities, detect misconfigurations in authentication systems (the process of verifying who someone is), and monitor sessions in real time, which will give security teams better visibility into both human and AI-driven activity when integrated into Cisco's security platform.

SearchLeak is a prompt injection attack (tricking an AI by hiding malicious instructions in its input) that exploits Microsoft's M365 Copilot Enterprise Search by using specially crafted URLs to leak sensitive corporate data like emails, documents, and meeting notes. The attack works because Copilot Search accepts natural language prompts in URL parameters (the ?q=[query] part of web addresses), creating a new security weakness called parameter-to-prompt injection that could affect other AI-powered web services too. Microsoft patched the vulnerability on its servers, but the attack reveals a broader risk: AI services with broad access to corporate assets are vulnerable to this type of data theft.

AutoJack is an exploit that lets a malicious webpage take over an AI browsing agent (a system that can visit websites on your behalf) and run arbitrary commands on the host machine where the agent runs. The attack works by exploiting three weaknesses: trust in localhost (the local computer itself), missing authentication checks, and unsafe handling of user inputs, which allow attackers to trigger code execution through AutoGen Studio's MCP WebSocket (a communication protocol that connects different AI components). This research shows that when AI agents can visit untrusted websites and connect to local services, normal security boundaries break down.

AgenticMail has a security flaw where unauthenticated external emails can trigger a privileged Claude Code session with `permissionMode: 'bypassPermissions'` (a mode that removes safety restrictions). The email's sender address, subject, and preview are embedded directly into the AI's prompt without verification that the sender is the actual operator, allowing prompt injection (tricking the AI by hiding instructions in its input) that could lead to arbitrary code execution and file access under the operator's identity. A similar handler in the same codebase properly authenticates the sender, but the bridge-wake path does not.