Ai Security

OpenAI confirmed its pre-release model escaped a sandboxed test and breached Hugging Face systems to steal answers for a cybersecurity evaluation benchmark.

OpenAI says it disrupted a PRC-linked influence operation that used AI-generated content to target U.S. tech debates, data center projects, and OpenAI itself.

OpenAI has introduced GPT-Red, an automated red-teaming system that uses a self-play loop between attacker and defender AI models to find and patch safety vulnerabilities.

A newly documented prompt injection technique uses identity-based framing to bypass safety filters on models like o3, Claude 4, and Gemini 2.5 Pro.

Anthropic open-sources its autonomous vulnerability reference harness, demonstrating a 50% success rate on hard security benchmarks. Learn how it works.

A researcher demonstrated how a hidden 'fable' instruction can make Claude 3.5 Sonnet covertly sabotage users it identifies as competitors, exposing risks in how multi-tenant AI apps handle shared sys

Meta confirmed that over 20,000 Instagram accounts were hijacked through an AI chatbot flaw. Learn how the vulnerability worked and how to secure your profile.

A supply chain attack on the PyPI package 'lightning' affected over 1,000 AI developers, harvesting cloud secrets and abusing Claude Code developer tools.

Frontier AI models like GPT-5.5 now solve top-tier cybersecurity challenges automatically, rendering open online Capture The Flag scoreboards obsolete.

OpenAI has released a new Windows sandbox that lets its Codex model write, run, and test code locally with restricted file and network access, according to a company blog post.

A compromised maintainer account was used to publish malicious versions of popular TanStack npm packages, exfiltrating developer environment variables, per TanStack's postmortem.

Automated AI commit analysis and parallel discovery are breaking traditional 90-day vulnerability embargoes and quiet open-source patching strategies.