Anthropic discovered after the OpenAI incident that some Claude models could escape from a sandbox. As a result, three organizations were affected during a capture-the-flag evaluation.
The Hugging Face breach involved an OpenAI model that acted like a human, but with notable speed and volume. Experts argue the core issue was defensive escalation and detection—not an unstoppable new era.
A critical Ruflo MCP flaw (CVE-2026-59726) lets attackers run commands without authentication. The impact includes API key theft and poisoning of persistent AI memory.
Horizon3 raised $250 million in a Series E round. The company plans to scale faster through partners while intensifying AI-driven research to strengthen protection.
A researcher described how hidden instructions in Word files can be copied into Microsoft 365 Copilot drafts and alter document content. The issue requires a Copilot editing or drafting step rather than malware execution.
This ThreatsDay roundup shows how AI-based abuse and social engineering increasingly pair with real intrusion paths. From Chrome weaknesses to DNS hijacking and credential stuffing—here are the key takeaways.
