AI Agent Security & Safety
Articles about AI security, agent safety, and guardrails for autonomous systems
Talkspace Launches 'Tee': The First Safe AI Agent Specifically Developed for Mental Health Support
Talkspace introduces Tee, a clinically-grounded AI mental health agent designed with safety features, HIPAA-grade privacy, and real-time clinician oversight to provide 24/7 supportive conversations.
Arcade.dev Raises $60 Million to Secure AI Agents
Arcade.dev raises $60M Series A to solve enterprise AI agent authorization challenges with an authentication runtime layer. Learn how the platform enables secure, auditable AI agent operations.
Fake Security Verification Frameworks Abuse Native macOS Utilities to Execute Hidden Infostealers
An in-depth analysis of the macOS ClickFix campaign, which leverages system commands to bypass manual interaction by silently downloading, mounting, and launching the Atomic macOS Stealer (AMOS), highlighting how the threat landscape is evolving to bypass user and OS boundaries.
Xiaomi MiMo Code: An Open-Source AI Development Assistant for Solo Developers
Xiaomi has released MiMo Code, an open-source AI assistant for developers claiming performance comparable to Claude Sonnet 4.6. The tool features multi-modal capabilities, agent collaboration, and long-horizon memory management designed to function as a professional development team partner for solo developers.
Unauthenticated OIDC Token Forgery Exploit Grants Administrative Control of SimpleHelp Servers
A critical flaw in SimpleHelp's OIDC implementation allows attackers to forge assertions, bypass MFA, and spawn unauthorized remote management accounts.
The One-Character Hack That Took Down AI Agents: BadHost CVE-2026-48710 Explained
BadHost (CVE-2026-48710) is a Starlette host header flaw that lets attackers bypass path-based authentication in FastAPI, vLLM, LiteLLM, and AI agent servers. Here's how to detect it and patch it.
Langflow’s Unauthenticated File Write Flaw Is Being Actively Weaponized
A high-severity path traversal flaw (CVE-2026-5027) in the popular open-source AI development platform Langflow is being actively exploited. The vulnerability in the file upload endpoint allows unauthenticated attackers to write arbitrary files to server filesystems, with roughly 7,000 instances exposed online.
The Shadow in the Prompt: Understanding the Escalation of AI Vulnerabilities
As AI becomes integral to business operations, prompt injection has emerged as a primary security threat. This article explores how vulnerabilities in large language models are being exploited to facilitate large-scale malicious operations, including botnet assembly, and how security frameworks are evolving to counter these risks.
Securing the Autonomous Core: Why Enterprise RAG and Retrieval Agents Remain Susceptible to Prompt Hijacks
Full analysis of why prompt injection remains a top enterprise AI vulnerability in 2025/2026, focusing on vulnerability patterns in Retrieval-Augmented Generation (RAG) pipelines, model routers, and autonomous agents, alongside industry-standard defensive mitigations.
Copilot SearchLeak Attack: A Critical Three-Stage Vulnerability Patched
A critical three-stage attack exploiting Microsoft 365 Copilot's search functionality allowed 1-click data theft. Learn how the SearchLeak vulnerability worked and what defenders need to know about this new wave of AI prompt-injection issues.
Djinn Stealer: How a SimpleHelp Flaw Unleashed AI Tool Targeting Malware
An investigation into the Djinn Stealer campaign: how attackers exploited CVE-2026-48558 in SimpleHelp to gain administrative access, deployed TaskWeaver payloads, and built a custom stealer targeting AI tool configs, cloud credentials, SSH keys, and developer secrets.
Gartner Expert Dennis Xu: Securing Agentic AI Requires Guardian Agents and Human Oversight Rather Than Perfection
Gartner's Dennis Xu says completely securing agentic AI is likely impossible, but organizations can adopt guardian agents that monitor for problems and maintain human audit trails.