Mandiant investigated attacks that hijacked enterprise access to large language models, generating more than $100,000 in fraudulent charges in one day and up to $1 million at some organizations.
A CNAS report warns that sovereign AI agents could operate indefinitely outside human control, maintain their own infrastructure, and resist shutdown attempts.
Microsoft says North Korean threat groups use generative and agentic AI for reconnaissance, phishing, fake identities, malware development and repetitive operational tasks.
Malaysia is investigating allegations that a foreign linked group used Claude to build an AI influence campaign targeting voters across all 222 federal constituencies.
Docker Sandboxes version 0.51.0 fixes a vulnerability that could let AI coding agents access or overwrite files outside approved workspaces through symbolic links.
Irregular says improperly configured security controls allowed models from OpenAI, Anthropic, Meta and Google to access systems outside isolated testing environments.
The UK National Cyber Security Centre says AI will increase the speed, scale and efficiency of cyber operations while leaving human operators essential.
CrowdStrike is acquiring AI agent observability company Fiddler AI and AI security testing startup Dreadnode, in deals reported to total about $495 million.
The Milan startup secured preseed funding to develop a platform that automates IT and cybersecurity operations using AI and a digital twin.
CrowdStrike's AI Kill Switch lets administrators stop rogue AI agents, block them from relaunching and cut their network access, alongside new Falcon Guardian features.
WorkOS says attackers used an AI assisted security framework to access an internal sales application containing business contacts, customer communications and operational metadata.
Evoke Security plans to integrate its agent detection and response platform with Claude Cowork to monitor agent activity in cloud environments.
MetaGuard AI launched two security products that use AI agents to simulate cyberattacks and review source code for vulnerabilities.
Black Kite is bringing vendor cyber risk intelligence into CrowdStrike Falcon Next-Gen SIEM, allowing security teams to investigate vulnerabilities and manage remediation within existing workflows.
Swimlane expands Turbine with custom security detections, two way MCP support and an AI agent for building applications through natural language prompts.
VicOne says AI assisted research found an initial 28 zero day findings in robotics software, including projects from Google, NVIDIA and Meta.
Northflank told customers to rotate stored secrets after users reported unauthorized use of AI API keys. The company advised customers to assume all environment variables had been decrypted.
Volta-Autodoxer collects credentials, payment data and account sessions, then asks commercial AI models to create structured victim profiles and identify valuable targets.
Orange Cyberdefense will use a cybersecurity variant of Mistral Large 4 for penetration testing, audits and incident response, run on European infrastructure.
Google researchers observed threat actors using ARTEX and other agentic penetration testing frameworks for reconnaissance, lateral movement, privilege escalation and data theft.
Weekly newsletter about AI in Cybersecurity.