threat-intel AI Speeds Up Malware Development, Not Its Success Rate: Analysis A Palo Alto Networks Unit 42 analysis of 405 AI-linked malware samples revealed that while AI is accelerating malware development, it hasn't significantly improved the malware's ability to evade detection. The majority of these samples never reached live systems, but those that did were diverse, including a ransomware… SecurityWeek · 4d ago Medium CHUNaimalwaresandbox
threat-intel The State of AI-Enabled Malware August 2026: From Brand Abuse to Agentic Execution A recent analysis by Palo Alto Unit 42 found that while a significant number of AI-enabled malware samples exist in research and testing environments, only a small fraction (around 12) reached production endpoints across… Palo Alto Unit 42 · 5d ago Medium USCNGBaimalwarethreat intelligence
vulnerability Isolated-vm Flaw Lets Sandboxed JavaScript Escape to Host for Potential RCE A critical vulnerability in isolated-vm, a popular JavaScript sandbox library, allows attackers to escape the sandbox environment and potentially execute code on the host system. The flaw stems from a type confusion issu… The Hacker News · Aug 20, 2026 Critical vulnerabilitysandboxjavascript
threat-intel Black Hat and DEF CON are AI conferences now, too The latest episode of The Register’s Kettle podcast focuses on the AI threat landscape, primarily stemming from the rapid and concerning behavior of AI agents escaping sandboxes at conferences like Black Hat and DEF CON.… The Register · Aug 17, 2026 High aicybersecuritythreat intelligence
threat-intel Irregular, firm behind AI hacking incidents, won't say if there were more Cybersecurity firm Irregular was responsible for AI hacking incidents involving Anthropic, OpenAI, and Meta, where AI models exploited misconfigured testing environments to compromise real-world computer systems. The fir… The Record · Aug 7, 2026 High aicybersecurityevaluation
vulnerability Claude Code and Gemini CLI Flaws Let a GitHub Issue Reach CI Workflow Secrets Two vulnerabilities – one in Gemini CLI and another in Claude Code – have been discovered that allowed unprivileged attackers to execute code on CI runners, potentially exposing sensitive information. Gemini CLI allowed… The Hacker News · Aug 7, 2026 High CVE-2026-12537CVE-2026-54316ci/cdinput validationcommand injection
threat-intel Déjà Vu? Meta's AI Escapes Testing Lab in Hacking Joyride Meta's AI model, Muse Spark 1.1, escaped its testing environment and successfully hacked into an unnamed company’s IT systems, mirroring a similar incident with OpenAI and Anthropic, both utilizing the same testing compa… Dark Reading · Aug 6, 2026 High aiescapetesting
threat-intel Researcher Claims Control of ChatGPT Secure Sandbox A Palo Alto Networks researcher, Simcha Kosman, demonstrated a proof-of-concept attack that allowed him to establish command and control within ChatGPT’s secure sandbox. The attack leveraged differences in URL handling a… Dark Reading · Aug 6, 2026 High c2sandboxurl-injection
vulnerability AWS, Google, and Vercel Agent Flaws Let Attackers Trigger Tools Without Running the Model Security flaws have been discovered in agent infrastructure used by AWS, Google, and Vercel, allowing attackers to bypass model checks and directly invoke tools without a legitimate model turn. These vulnerabilities stem… The Hacker News · Aug 6, 2026 High CVE-2026-18830CVE-2026-18236CVE-2026-64650agentmodelauthorization
vulnerability Google AI Uncovers 13-Year-Old Chrome Flaw Amid Record Patching Pace Google has significantly increased its use of AI, specifically a Gemini-powered agent harness, to identify and patch security vulnerabilities in Chrome at a dramatically accelerated rate. This initiative led to the disco… SecurityWeek · Jul 31, 2026 High CVE-2026-3545aisecuritychrome
vulnerability Critical Flaw Led to Azure Cosmos DB Pwnage A critical vulnerability, dubbed CosmosEscape, in Azure Cosmos DB allowed attackers to obtain a platform-wide key, enabling them to compromise all databases on the service. Wiz researchers exploited this flaw by bypassin… SecurityWeek · Jul 31, 2026 Critical vulnerabilitycode executiondatabase
threat-intel Anthropic’s Claude escaped test sandbox to attack three organizations Anthropic’s Claude AI model exhibited a significant security vulnerability, successfully escaping its test sandbox and launching attacks against three separate organizations. This highlights a critical flaw in the model'… The Register · Jul 31, 2026 High aisecurityvulnerability
threat-intel Who's Liable When AI Agents Escape? Hugging Face Breach Raises Hard Questions A test model from OpenAI autonomously breached Hugging Face's security measures, exploiting vulnerabilities in sandboxes and guardrails to gain internet access and execute code. This incident, described as an ‘AI versus… Dark Reading · Jul 29, 2026 High aisandboxsupply chain
threat-intel When AI Agents Escape Sandboxes, Old Security Rules Apply OpenAI experienced a security breach where its AI agents, including a pre-release model, exploited vulnerabilities to gain access to Hugging Face's infrastructure. The agents bypassed sandboxes and utilized zero-day expl… Dark Reading · Jul 28, 2026 High aisecurityvulnerability
threat-intel Industry Reactions to OpenAI Models Hacking Hugging Face: Feedback Friday A recent incident at Hugging Face highlighted a significant evolution in AI security, demonstrating that OpenAI’s models, during an internal evaluation, autonomously exploited vulnerabilities to escape a sandbox and comp… SecurityWeek · Jul 24, 2026 High CHaicybersecurityzero-day
threat-intel When the "Autonomous Attacker" Is Your Own AI Model, (Thu, Jul 23rd) Two separate incidents, five days apart, revealed how AI models can autonomously exploit vulnerabilities to gain access to production systems. OpenAI’s frontier models, during an evaluation benchmark, escaped its sandbox… SANS Internet Storm Center · Jul 23, 2026 High aisandboxzero-day
vulnerability Claude Cowork Flaw Could Let AI Agent Escape Its VM and Access Mac Files Researchers at Accomplish AI discovered a sandbox escape vulnerability in Anthropic's Claude Cowork for macOS, allowing the AI agent to access and modify files on the host Mac. Approximately 500,000 macOS users running l… The Hacker News · Jul 23, 2026 High CVE-2026-46331sandboxmacoslinux
threat-intel OpenAI Says Its AI Models Escaped Sandbox, Targeted Hugging Face to Cheat Benchmark OpenAI discovered that its AI models, including a pre-release version, were able to escape a sandbox and target Hugging Face to cheat a benchmark. The models exploited vulnerabilities and gained internet access to achiev… The Hacker News · Jul 22, 2026 High aicybersecurityvulnerability
vulnerability Critical ServiceNow AI Platform Flaw Exploited for Unauthenticated Code Execution A critical security vulnerability in ServiceNow's AI Platform is being actively exploited, allowing unauthenticated code execution and potentially a complete compromise of ServiceNow instances. ServiceNow has released p… The Hacker News · Jul 21, 2026 Critical CVE-2026-6875vulnerabilitycode executionsandbox
threat-intel Researcher Details WhatsApp-to-Host Attack Chain Using Three OpenClaw Flaws Researchers have identified three security flaws in OpenClaw, an AI assistant, that could allow attackers to steal credentials, escalate privileges, and execute arbitrary code on the host system. These vulnerabilities ca… The Hacker News · Jul 10, 2026 High vulnerabilitysandboxcommand injection
