AI security

Check Point and BeyondTrust disclose two distinct vulnerabilities in OpenAI products that allowed covert data theft through DNS queries and GitHub token compromise via branch name injection.

Security researchers disclose a critical deserialization flaw and two high-severity bugs in the most widely downloaded AI framework on PyPI, with patches now available.

Threat actor TeamPCP used credentials stolen in the Trivy compromise to backdoor LiteLLM versions 1.82.7 and 1.82.8 on PyPI, deploying a multi-stage credential stealer across an estimated 500,000 environments.

Anthropic, AWS, GitHub, Google, Google DeepMind, Microsoft, and OpenAI fund a Linux Foundation initiative to help overwhelmed open source maintainers triage AI-generated vulnerability reports.

The cybersecurity industry's largest conference was dominated by agentic AI security products, even as data shows only 5 percent of enterprises have moved AI agents into production.

CVE-2026-33017, a CVSS 9.3 unauthenticated RCE in the open-source AI framework Langflow, was weaponized within 20 hours of its March 17 advisory with no public exploit code available.

Microsoft, the Pentagon, and Zscaler are extending zero trust beyond network perimeters to AI agents, weapons systems, and sovereign data flows, driven by a non-human identity explosion that outnumbers human users 100 to 1.

IBM's annual threat index finds vulnerability exploitation now causes 40% of breaches, with 109 ransomware groups active and over 300,000 AI platform credentials stolen.